If consumed does not match nalsize, favor nalsize.
[ffmpeg.git] / libavcodec / h264.c
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file h264.c
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #include "dsputil.h"
29 #include "avcodec.h"
30 #include "mpegvideo.h"
31 #include "h264.h"
32 #include "h264data.h"
33 #include "h264_parser.h"
34 #include "golomb.h"
35 #include "rectangle.h"
36
37 #include "cabac.h"
38
39 //#undef NDEBUG
40 #include <assert.h>
41
42 /**
43  * Value of Picture.reference when Picture is not a reference picture, but
44  * is held for delayed output.
45  */
46 #define DELAYED_PIC_REF 4
47
48 static VLC coeff_token_vlc[4];
49 static VLC chroma_dc_coeff_token_vlc;
50
51 static VLC total_zeros_vlc[15];
52 static VLC chroma_dc_total_zeros_vlc[3];
53
54 static VLC run_vlc[6];
55 static VLC run7_vlc;
56
57 static void svq3_luma_dc_dequant_idct_c(DCTELEM *block, int qp);
58 static void svq3_add_idct_c(uint8_t *dst, DCTELEM *block, int stride, int qp, int dc);
59 static void filter_mb( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
60 static void filter_mb_fast( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
61
62 static av_always_inline uint32_t pack16to32(int a, int b){
63 #ifdef WORDS_BIGENDIAN
64    return (b&0xFFFF) + (a<<16);
65 #else
66    return (a&0xFFFF) + (b<<16);
67 #endif
68 }
69
70 const uint8_t ff_rem6[52]={
71 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
72 };
73
74 const uint8_t ff_div6[52]={
75 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8,
76 };
77
78
79 static void fill_caches(H264Context *h, int mb_type, int for_deblock){
80     MpegEncContext * const s = &h->s;
81     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
82     int topleft_xy, top_xy, topright_xy, left_xy[2];
83     int topleft_type, top_type, topright_type, left_type[2];
84     int left_block[8];
85     int topleft_partition= -1;
86     int i;
87
88     top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);
89
90     //FIXME deblocking could skip the intra and nnz parts.
91     if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[top_xy]) && !FRAME_MBAFF)
92         return;
93
94     /* Wow, what a mess, why didn't they simplify the interlacing & intra
95      * stuff, I can't imagine that these complex rules are worth it. */
96
97     topleft_xy = top_xy - 1;
98     topright_xy= top_xy + 1;
99     left_xy[1] = left_xy[0] = mb_xy-1;
100     left_block[0]= 0;
101     left_block[1]= 1;
102     left_block[2]= 2;
103     left_block[3]= 3;
104     left_block[4]= 7;
105     left_block[5]= 10;
106     left_block[6]= 8;
107     left_block[7]= 11;
108     if(FRAME_MBAFF){
109         const int pair_xy          = s->mb_x     + (s->mb_y & ~1)*s->mb_stride;
110         const int top_pair_xy      = pair_xy     - s->mb_stride;
111         const int topleft_pair_xy  = top_pair_xy - 1;
112         const int topright_pair_xy = top_pair_xy + 1;
113         const int topleft_mb_frame_flag  = !IS_INTERLACED(s->current_picture.mb_type[topleft_pair_xy]);
114         const int top_mb_frame_flag      = !IS_INTERLACED(s->current_picture.mb_type[top_pair_xy]);
115         const int topright_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[topright_pair_xy]);
116         const int left_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[pair_xy-1]);
117         const int curr_mb_frame_flag = !IS_INTERLACED(mb_type);
118         const int bottom = (s->mb_y & 1);
119         tprintf(s->avctx, "fill_caches: curr_mb_frame_flag:%d, left_mb_frame_flag:%d, topleft_mb_frame_flag:%d, top_mb_frame_flag:%d, topright_mb_frame_flag:%d\n", curr_mb_frame_flag, left_mb_frame_flag, topleft_mb_frame_flag, top_mb_frame_flag, topright_mb_frame_flag);
120         if (bottom
121                 ? !curr_mb_frame_flag // bottom macroblock
122                 : (!curr_mb_frame_flag && !top_mb_frame_flag) // top macroblock
123                 ) {
124             top_xy -= s->mb_stride;
125         }
126         if (bottom
127                 ? !curr_mb_frame_flag // bottom macroblock
128                 : (!curr_mb_frame_flag && !topleft_mb_frame_flag) // top macroblock
129                 ) {
130             topleft_xy -= s->mb_stride;
131         } else if(bottom && curr_mb_frame_flag && !left_mb_frame_flag) {
132             topleft_xy += s->mb_stride;
133             // take topleft mv from the middle of the mb, as opposed to all other modes which use the bottom-right partition
134             topleft_partition = 0;
135         }
136         if (bottom
137                 ? !curr_mb_frame_flag // bottom macroblock
138                 : (!curr_mb_frame_flag && !topright_mb_frame_flag) // top macroblock
139                 ) {
140             topright_xy -= s->mb_stride;
141         }
142         if (left_mb_frame_flag != curr_mb_frame_flag) {
143             left_xy[1] = left_xy[0] = pair_xy - 1;
144             if (curr_mb_frame_flag) {
145                 if (bottom) {
146                     left_block[0]= 2;
147                     left_block[1]= 2;
148                     left_block[2]= 3;
149                     left_block[3]= 3;
150                     left_block[4]= 8;
151                     left_block[5]= 11;
152                     left_block[6]= 8;
153                     left_block[7]= 11;
154                 } else {
155                     left_block[0]= 0;
156                     left_block[1]= 0;
157                     left_block[2]= 1;
158                     left_block[3]= 1;
159                     left_block[4]= 7;
160                     left_block[5]= 10;
161                     left_block[6]= 7;
162                     left_block[7]= 10;
163                 }
164             } else {
165                 left_xy[1] += s->mb_stride;
166                 //left_block[0]= 0;
167                 left_block[1]= 2;
168                 left_block[2]= 0;
169                 left_block[3]= 2;
170                 //left_block[4]= 7;
171                 left_block[5]= 10;
172                 left_block[6]= 7;
173                 left_block[7]= 10;
174             }
175         }
176     }
177
178     h->top_mb_xy = top_xy;
179     h->left_mb_xy[0] = left_xy[0];
180     h->left_mb_xy[1] = left_xy[1];
181     if(for_deblock){
182         topleft_type = 0;
183         topright_type = 0;
184         top_type     = h->slice_table[top_xy     ] < 255 ? s->current_picture.mb_type[top_xy]     : 0;
185         left_type[0] = h->slice_table[left_xy[0] ] < 255 ? s->current_picture.mb_type[left_xy[0]] : 0;
186         left_type[1] = h->slice_table[left_xy[1] ] < 255 ? s->current_picture.mb_type[left_xy[1]] : 0;
187
188         if(FRAME_MBAFF && !IS_INTRA(mb_type)){
189             int list;
190             int v = *(uint16_t*)&h->non_zero_count[mb_xy][14];
191             for(i=0; i<16; i++)
192                 h->non_zero_count_cache[scan8[i]] = (v>>i)&1;
193             for(list=0; list<h->list_count; list++){
194                 if(USES_LIST(mb_type,list)){
195                     uint32_t *src = (uint32_t*)s->current_picture.motion_val[list][h->mb2b_xy[mb_xy]];
196                     uint32_t *dst = (uint32_t*)h->mv_cache[list][scan8[0]];
197                     int8_t *ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
198                     for(i=0; i<4; i++, dst+=8, src+=h->b_stride){
199                         dst[0] = src[0];
200                         dst[1] = src[1];
201                         dst[2] = src[2];
202                         dst[3] = src[3];
203                     }
204                     *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
205                     *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = pack16to32(ref[0],ref[1])*0x0101;
206                     ref += h->b8_stride;
207                     *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
208                     *(uint32_t*)&h->ref_cache[list][scan8[10]] = pack16to32(ref[0],ref[1])*0x0101;
209                 }else{
210                     fill_rectangle(&h-> mv_cache[list][scan8[ 0]], 4, 4, 8, 0, 4);
211                     fill_rectangle(&h->ref_cache[list][scan8[ 0]], 4, 4, 8, (uint8_t)LIST_NOT_USED, 1);
212                 }
213             }
214         }
215     }else{
216         topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
217         top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
218         topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
219         left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
220         left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
221     }
222
223     if(IS_INTRA(mb_type)){
224         h->topleft_samples_available=
225         h->top_samples_available=
226         h->left_samples_available= 0xFFFF;
227         h->topright_samples_available= 0xEEEA;
228
229         if(!IS_INTRA(top_type) && (top_type==0 || h->pps.constrained_intra_pred)){
230             h->topleft_samples_available= 0xB3FF;
231             h->top_samples_available= 0x33FF;
232             h->topright_samples_available= 0x26EA;
233         }
234         for(i=0; i<2; i++){
235             if(!IS_INTRA(left_type[i]) && (left_type[i]==0 || h->pps.constrained_intra_pred)){
236                 h->topleft_samples_available&= 0xDF5F;
237                 h->left_samples_available&= 0x5F5F;
238             }
239         }
240
241         if(!IS_INTRA(topleft_type) && (topleft_type==0 || h->pps.constrained_intra_pred))
242             h->topleft_samples_available&= 0x7FFF;
243
244         if(!IS_INTRA(topright_type) && (topright_type==0 || h->pps.constrained_intra_pred))
245             h->topright_samples_available&= 0xFBFF;
246
247         if(IS_INTRA4x4(mb_type)){
248             if(IS_INTRA4x4(top_type)){
249                 h->intra4x4_pred_mode_cache[4+8*0]= h->intra4x4_pred_mode[top_xy][4];
250                 h->intra4x4_pred_mode_cache[5+8*0]= h->intra4x4_pred_mode[top_xy][5];
251                 h->intra4x4_pred_mode_cache[6+8*0]= h->intra4x4_pred_mode[top_xy][6];
252                 h->intra4x4_pred_mode_cache[7+8*0]= h->intra4x4_pred_mode[top_xy][3];
253             }else{
254                 int pred;
255                 if(!top_type || (IS_INTER(top_type) && h->pps.constrained_intra_pred))
256                     pred= -1;
257                 else{
258                     pred= 2;
259                 }
260                 h->intra4x4_pred_mode_cache[4+8*0]=
261                 h->intra4x4_pred_mode_cache[5+8*0]=
262                 h->intra4x4_pred_mode_cache[6+8*0]=
263                 h->intra4x4_pred_mode_cache[7+8*0]= pred;
264             }
265             for(i=0; i<2; i++){
266                 if(IS_INTRA4x4(left_type[i])){
267                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[0+2*i]];
268                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[1+2*i]];
269                 }else{
270                     int pred;
271                     if(!left_type[i] || (IS_INTER(left_type[i]) && h->pps.constrained_intra_pred))
272                         pred= -1;
273                     else{
274                         pred= 2;
275                     }
276                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]=
277                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= pred;
278                 }
279             }
280         }
281     }
282
283
284 /*
285 0 . T T. T T T T
286 1 L . .L . . . .
287 2 L . .L . . . .
288 3 . T TL . . . .
289 4 L . .L . . . .
290 5 L . .. . . . .
291 */
292 //FIXME constraint_intra_pred & partitioning & nnz (lets hope this is just a typo in the spec)
293     if(top_type){
294         h->non_zero_count_cache[4+8*0]= h->non_zero_count[top_xy][4];
295         h->non_zero_count_cache[5+8*0]= h->non_zero_count[top_xy][5];
296         h->non_zero_count_cache[6+8*0]= h->non_zero_count[top_xy][6];
297         h->non_zero_count_cache[7+8*0]= h->non_zero_count[top_xy][3];
298
299         h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][9];
300         h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][8];
301
302         h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][12];
303         h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][11];
304
305     }else{
306         h->non_zero_count_cache[4+8*0]=
307         h->non_zero_count_cache[5+8*0]=
308         h->non_zero_count_cache[6+8*0]=
309         h->non_zero_count_cache[7+8*0]=
310
311         h->non_zero_count_cache[1+8*0]=
312         h->non_zero_count_cache[2+8*0]=
313
314         h->non_zero_count_cache[1+8*3]=
315         h->non_zero_count_cache[2+8*3]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
316
317     }
318
319     for (i=0; i<2; i++) {
320         if(left_type[i]){
321             h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[0+2*i]];
322             h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[1+2*i]];
323             h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[4+2*i]];
324             h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[5+2*i]];
325         }else{
326             h->non_zero_count_cache[3+8*1 + 2*8*i]=
327             h->non_zero_count_cache[3+8*2 + 2*8*i]=
328             h->non_zero_count_cache[0+8*1 +   8*i]=
329             h->non_zero_count_cache[0+8*4 +   8*i]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
330         }
331     }
332
333     if( h->pps.cabac ) {
334         // top_cbp
335         if(top_type) {
336             h->top_cbp = h->cbp_table[top_xy];
337         } else if(IS_INTRA(mb_type)) {
338             h->top_cbp = 0x1C0;
339         } else {
340             h->top_cbp = 0;
341         }
342         // left_cbp
343         if (left_type[0]) {
344             h->left_cbp = h->cbp_table[left_xy[0]] & 0x1f0;
345         } else if(IS_INTRA(mb_type)) {
346             h->left_cbp = 0x1C0;
347         } else {
348             h->left_cbp = 0;
349         }
350         if (left_type[0]) {
351             h->left_cbp |= ((h->cbp_table[left_xy[0]]>>((left_block[0]&(~1))+1))&0x1) << 1;
352         }
353         if (left_type[1]) {
354             h->left_cbp |= ((h->cbp_table[left_xy[1]]>>((left_block[2]&(~1))+1))&0x1) << 3;
355         }
356     }
357
358 #if 1
359     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
360         int list;
361         for(list=0; list<h->list_count; list++){
362             if(!USES_LIST(mb_type, list) && !IS_DIRECT(mb_type) && !h->deblocking_filter){
363                 /*if(!h->mv_cache_clean[list]){
364                     memset(h->mv_cache [list],  0, 8*5*2*sizeof(int16_t)); //FIXME clean only input? clean at all?
365                     memset(h->ref_cache[list], PART_NOT_AVAILABLE, 8*5*sizeof(int8_t));
366                     h->mv_cache_clean[list]= 1;
367                 }*/
368                 continue;
369             }
370             h->mv_cache_clean[list]= 0;
371
372             if(USES_LIST(top_type, list)){
373                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
374                 const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
375                 *(uint32_t*)h->mv_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 0];
376                 *(uint32_t*)h->mv_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 1];
377                 *(uint32_t*)h->mv_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 2];
378                 *(uint32_t*)h->mv_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 3];
379                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
380                 h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
381                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
382                 h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
383             }else{
384                 *(uint32_t*)h->mv_cache [list][scan8[0] + 0 - 1*8]=
385                 *(uint32_t*)h->mv_cache [list][scan8[0] + 1 - 1*8]=
386                 *(uint32_t*)h->mv_cache [list][scan8[0] + 2 - 1*8]=
387                 *(uint32_t*)h->mv_cache [list][scan8[0] + 3 - 1*8]= 0;
388                 *(uint32_t*)&h->ref_cache[list][scan8[0] + 0 - 1*8]= ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101;
389             }
390
391             for(i=0; i<2; i++){
392                 int cache_idx = scan8[0] - 1 + i*2*8;
393                 if(USES_LIST(left_type[i], list)){
394                     const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
395                     const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
396                     *(uint32_t*)h->mv_cache[list][cache_idx  ]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[0+i*2]];
397                     *(uint32_t*)h->mv_cache[list][cache_idx+8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[1+i*2]];
398                     h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
399                     h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
400                 }else{
401                     *(uint32_t*)h->mv_cache [list][cache_idx  ]=
402                     *(uint32_t*)h->mv_cache [list][cache_idx+8]= 0;
403                     h->ref_cache[list][cache_idx  ]=
404                     h->ref_cache[list][cache_idx+8]= left_type[i] ? LIST_NOT_USED : PART_NOT_AVAILABLE;
405                 }
406             }
407
408             if((for_deblock || (IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred)) && !FRAME_MBAFF)
409                 continue;
410
411             if(USES_LIST(topleft_type, list)){
412                 const int b_xy = h->mb2b_xy[topleft_xy] + 3 + h->b_stride + (topleft_partition & 2*h->b_stride);
413                 const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + (topleft_partition & h->b8_stride);
414                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
415                 h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
416             }else{
417                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= 0;
418                 h->ref_cache[list][scan8[0] - 1 - 1*8]= topleft_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
419             }
420
421             if(USES_LIST(topright_type, list)){
422                 const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
423                 const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
424                 *(uint32_t*)h->mv_cache[list][scan8[0] + 4 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
425                 h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
426             }else{
427                 *(uint32_t*)h->mv_cache [list][scan8[0] + 4 - 1*8]= 0;
428                 h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
429             }
430
431             if((IS_SKIP(mb_type) || IS_DIRECT(mb_type)) && !FRAME_MBAFF)
432                 continue;
433
434             h->ref_cache[list][scan8[5 ]+1] =
435             h->ref_cache[list][scan8[7 ]+1] =
436             h->ref_cache[list][scan8[13]+1] =  //FIXME remove past 3 (init somewhere else)
437             h->ref_cache[list][scan8[4 ]] =
438             h->ref_cache[list][scan8[12]] = PART_NOT_AVAILABLE;
439             *(uint32_t*)h->mv_cache [list][scan8[5 ]+1]=
440             *(uint32_t*)h->mv_cache [list][scan8[7 ]+1]=
441             *(uint32_t*)h->mv_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
442             *(uint32_t*)h->mv_cache [list][scan8[4 ]]=
443             *(uint32_t*)h->mv_cache [list][scan8[12]]= 0;
444
445             if( h->pps.cabac ) {
446                 /* XXX beurk, Load mvd */
447                 if(USES_LIST(top_type, list)){
448                     const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
449                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 0];
450                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 1];
451                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 2];
452                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 3];
453                 }else{
454                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 0 - 1*8]=
455                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 1 - 1*8]=
456                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 2 - 1*8]=
457                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 3 - 1*8]= 0;
458                 }
459                 if(USES_LIST(left_type[0], list)){
460                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
461                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 0*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[0]];
462                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[1]];
463                 }else{
464                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 0*8]=
465                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 1*8]= 0;
466                 }
467                 if(USES_LIST(left_type[1], list)){
468                     const int b_xy= h->mb2b_xy[left_xy[1]] + 3;
469                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 2*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[2]];
470                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 3*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[3]];
471                 }else{
472                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 2*8]=
473                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 3*8]= 0;
474                 }
475                 *(uint32_t*)h->mvd_cache [list][scan8[5 ]+1]=
476                 *(uint32_t*)h->mvd_cache [list][scan8[7 ]+1]=
477                 *(uint32_t*)h->mvd_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
478                 *(uint32_t*)h->mvd_cache [list][scan8[4 ]]=
479                 *(uint32_t*)h->mvd_cache [list][scan8[12]]= 0;
480
481                 if(h->slice_type == FF_B_TYPE){
482                     fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, 0, 1);
483
484                     if(IS_DIRECT(top_type)){
485                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0x01010101;
486                     }else if(IS_8X8(top_type)){
487                         int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride;
488                         h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy];
489                         h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1];
490                     }else{
491                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0;
492                     }
493
494                     if(IS_DIRECT(left_type[0]))
495                         h->direct_cache[scan8[0] - 1 + 0*8]= 1;
496                     else if(IS_8X8(left_type[0]))
497                         h->direct_cache[scan8[0] - 1 + 0*8]= h->direct_table[h->mb2b8_xy[left_xy[0]] + 1 + h->b8_stride*(left_block[0]>>1)];
498                     else
499                         h->direct_cache[scan8[0] - 1 + 0*8]= 0;
500
501                     if(IS_DIRECT(left_type[1]))
502                         h->direct_cache[scan8[0] - 1 + 2*8]= 1;
503                     else if(IS_8X8(left_type[1]))
504                         h->direct_cache[scan8[0] - 1 + 2*8]= h->direct_table[h->mb2b8_xy[left_xy[1]] + 1 + h->b8_stride*(left_block[2]>>1)];
505                     else
506                         h->direct_cache[scan8[0] - 1 + 2*8]= 0;
507                 }
508             }
509
510             if(FRAME_MBAFF){
511 #define MAP_MVS\
512                     MAP_F2F(scan8[0] - 1 - 1*8, topleft_type)\
513                     MAP_F2F(scan8[0] + 0 - 1*8, top_type)\
514                     MAP_F2F(scan8[0] + 1 - 1*8, top_type)\
515                     MAP_F2F(scan8[0] + 2 - 1*8, top_type)\
516                     MAP_F2F(scan8[0] + 3 - 1*8, top_type)\
517                     MAP_F2F(scan8[0] + 4 - 1*8, topright_type)\
518                     MAP_F2F(scan8[0] - 1 + 0*8, left_type[0])\
519                     MAP_F2F(scan8[0] - 1 + 1*8, left_type[0])\
520                     MAP_F2F(scan8[0] - 1 + 2*8, left_type[1])\
521                     MAP_F2F(scan8[0] - 1 + 3*8, left_type[1])
522                 if(MB_FIELD){
523 #define MAP_F2F(idx, mb_type)\
524                     if(!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
525                         h->ref_cache[list][idx] <<= 1;\
526                         h->mv_cache[list][idx][1] /= 2;\
527                         h->mvd_cache[list][idx][1] /= 2;\
528                     }
529                     MAP_MVS
530 #undef MAP_F2F
531                 }else{
532 #define MAP_F2F(idx, mb_type)\
533                     if(IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
534                         h->ref_cache[list][idx] >>= 1;\
535                         h->mv_cache[list][idx][1] <<= 1;\
536                         h->mvd_cache[list][idx][1] <<= 1;\
537                     }
538                     MAP_MVS
539 #undef MAP_F2F
540                 }
541             }
542         }
543     }
544 #endif
545
546     h->neighbor_transform_size= !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[0]);
547 }
548
549 static inline void write_back_intra_pred_mode(H264Context *h){
550     MpegEncContext * const s = &h->s;
551     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
552
553     h->intra4x4_pred_mode[mb_xy][0]= h->intra4x4_pred_mode_cache[7+8*1];
554     h->intra4x4_pred_mode[mb_xy][1]= h->intra4x4_pred_mode_cache[7+8*2];
555     h->intra4x4_pred_mode[mb_xy][2]= h->intra4x4_pred_mode_cache[7+8*3];
556     h->intra4x4_pred_mode[mb_xy][3]= h->intra4x4_pred_mode_cache[7+8*4];
557     h->intra4x4_pred_mode[mb_xy][4]= h->intra4x4_pred_mode_cache[4+8*4];
558     h->intra4x4_pred_mode[mb_xy][5]= h->intra4x4_pred_mode_cache[5+8*4];
559     h->intra4x4_pred_mode[mb_xy][6]= h->intra4x4_pred_mode_cache[6+8*4];
560 }
561
562 /**
563  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
564  */
565 static inline int check_intra4x4_pred_mode(H264Context *h){
566     MpegEncContext * const s = &h->s;
567     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
568     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
569     int i;
570
571     if(!(h->top_samples_available&0x8000)){
572         for(i=0; i<4; i++){
573             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
574             if(status<0){
575                 av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
576                 return -1;
577             } else if(status){
578                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
579             }
580         }
581     }
582
583     if(!(h->left_samples_available&0x8000)){
584         for(i=0; i<4; i++){
585             int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
586             if(status<0){
587                 av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
588                 return -1;
589             } else if(status){
590                 h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
591             }
592         }
593     }
594
595     return 0;
596 } //FIXME cleanup like next
597
598 /**
599  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
600  */
601 static inline int check_intra_pred_mode(H264Context *h, int mode){
602     MpegEncContext * const s = &h->s;
603     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
604     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
605
606     if(mode > 6U) {
607         av_log(h->s.avctx, AV_LOG_ERROR, "out of range intra chroma pred mode at %d %d\n", s->mb_x, s->mb_y);
608         return -1;
609     }
610
611     if(!(h->top_samples_available&0x8000)){
612         mode= top[ mode ];
613         if(mode<0){
614             av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
615             return -1;
616         }
617     }
618
619     if(!(h->left_samples_available&0x8000)){
620         mode= left[ mode ];
621         if(mode<0){
622             av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
623             return -1;
624         }
625     }
626
627     return mode;
628 }
629
630 /**
631  * gets the predicted intra4x4 prediction mode.
632  */
633 static inline int pred_intra_mode(H264Context *h, int n){
634     const int index8= scan8[n];
635     const int left= h->intra4x4_pred_mode_cache[index8 - 1];
636     const int top = h->intra4x4_pred_mode_cache[index8 - 8];
637     const int min= FFMIN(left, top);
638
639     tprintf(h->s.avctx, "mode:%d %d min:%d\n", left ,top, min);
640
641     if(min<0) return DC_PRED;
642     else      return min;
643 }
644
645 static inline void write_back_non_zero_count(H264Context *h){
646     MpegEncContext * const s = &h->s;
647     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
648
649     h->non_zero_count[mb_xy][0]= h->non_zero_count_cache[7+8*1];
650     h->non_zero_count[mb_xy][1]= h->non_zero_count_cache[7+8*2];
651     h->non_zero_count[mb_xy][2]= h->non_zero_count_cache[7+8*3];
652     h->non_zero_count[mb_xy][3]= h->non_zero_count_cache[7+8*4];
653     h->non_zero_count[mb_xy][4]= h->non_zero_count_cache[4+8*4];
654     h->non_zero_count[mb_xy][5]= h->non_zero_count_cache[5+8*4];
655     h->non_zero_count[mb_xy][6]= h->non_zero_count_cache[6+8*4];
656
657     h->non_zero_count[mb_xy][9]= h->non_zero_count_cache[1+8*2];
658     h->non_zero_count[mb_xy][8]= h->non_zero_count_cache[2+8*2];
659     h->non_zero_count[mb_xy][7]= h->non_zero_count_cache[2+8*1];
660
661     h->non_zero_count[mb_xy][12]=h->non_zero_count_cache[1+8*5];
662     h->non_zero_count[mb_xy][11]=h->non_zero_count_cache[2+8*5];
663     h->non_zero_count[mb_xy][10]=h->non_zero_count_cache[2+8*4];
664
665     if(FRAME_MBAFF){
666         // store all luma nnzs, for deblocking
667         int v = 0, i;
668         for(i=0; i<16; i++)
669             v += (!!h->non_zero_count_cache[scan8[i]]) << i;
670         *(uint16_t*)&h->non_zero_count[mb_xy][14] = v;
671     }
672 }
673
674 /**
675  * gets the predicted number of non zero coefficients.
676  * @param n block index
677  */
678 static inline int pred_non_zero_count(H264Context *h, int n){
679     const int index8= scan8[n];
680     const int left= h->non_zero_count_cache[index8 - 1];
681     const int top = h->non_zero_count_cache[index8 - 8];
682     int i= left + top;
683
684     if(i<64) i= (i+1)>>1;
685
686     tprintf(h->s.avctx, "pred_nnz L%X T%X n%d s%d P%X\n", left, top, n, scan8[n], i&31);
687
688     return i&31;
689 }
690
691 static inline int fetch_diagonal_mv(H264Context *h, const int16_t **C, int i, int list, int part_width){
692     const int topright_ref= h->ref_cache[list][ i - 8 + part_width ];
693     MpegEncContext *s = &h->s;
694
695     /* there is no consistent mapping of mvs to neighboring locations that will
696      * make mbaff happy, so we can't move all this logic to fill_caches */
697     if(FRAME_MBAFF){
698         const uint32_t *mb_types = s->current_picture_ptr->mb_type;
699         const int16_t *mv;
700         *(uint32_t*)h->mv_cache[list][scan8[0]-2] = 0;
701         *C = h->mv_cache[list][scan8[0]-2];
702
703         if(!MB_FIELD
704            && (s->mb_y&1) && i < scan8[0]+8 && topright_ref != PART_NOT_AVAILABLE){
705             int topright_xy = s->mb_x + (s->mb_y-1)*s->mb_stride + (i == scan8[0]+3);
706             if(IS_INTERLACED(mb_types[topright_xy])){
707 #define SET_DIAG_MV(MV_OP, REF_OP, X4, Y4)\
708                 const int x4 = X4, y4 = Y4;\
709                 const int mb_type = mb_types[(x4>>2)+(y4>>2)*s->mb_stride];\
710                 if(!USES_LIST(mb_type,list))\
711                     return LIST_NOT_USED;\
712                 mv = s->current_picture_ptr->motion_val[list][x4 + y4*h->b_stride];\
713                 h->mv_cache[list][scan8[0]-2][0] = mv[0];\
714                 h->mv_cache[list][scan8[0]-2][1] = mv[1] MV_OP;\
715                 return s->current_picture_ptr->ref_index[list][(x4>>1) + (y4>>1)*h->b8_stride] REF_OP;
716
717                 SET_DIAG_MV(*2, >>1, s->mb_x*4+(i&7)-4+part_width, s->mb_y*4-1);
718             }
719         }
720         if(topright_ref == PART_NOT_AVAILABLE
721            && ((s->mb_y&1) || i >= scan8[0]+8) && (i&7)==4
722            && h->ref_cache[list][scan8[0]-1] != PART_NOT_AVAILABLE){
723             if(!MB_FIELD
724                && IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
725                 SET_DIAG_MV(*2, >>1, s->mb_x*4-1, (s->mb_y|1)*4+(s->mb_y&1)*2+(i>>4)-1);
726             }
727             if(MB_FIELD
728                && !IS_INTERLACED(mb_types[h->left_mb_xy[0]])
729                && i >= scan8[0]+8){
730                 // leftshift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's ok.
731                 SET_DIAG_MV(/2, <<1, s->mb_x*4-1, (s->mb_y&~1)*4 - 1 + ((i-scan8[0])>>3)*2);
732             }
733         }
734 #undef SET_DIAG_MV
735     }
736
737     if(topright_ref != PART_NOT_AVAILABLE){
738         *C= h->mv_cache[list][ i - 8 + part_width ];
739         return topright_ref;
740     }else{
741         tprintf(s->avctx, "topright MV not available\n");
742
743         *C= h->mv_cache[list][ i - 8 - 1 ];
744         return h->ref_cache[list][ i - 8 - 1 ];
745     }
746 }
747
748 /**
749  * gets the predicted MV.
750  * @param n the block index
751  * @param part_width the width of the partition (4, 8,16) -> (1, 2, 4)
752  * @param mx the x component of the predicted motion vector
753  * @param my the y component of the predicted motion vector
754  */
755 static inline void pred_motion(H264Context * const h, int n, int part_width, int list, int ref, int * const mx, int * const my){
756     const int index8= scan8[n];
757     const int top_ref=      h->ref_cache[list][ index8 - 8 ];
758     const int left_ref=     h->ref_cache[list][ index8 - 1 ];
759     const int16_t * const A= h->mv_cache[list][ index8 - 1 ];
760     const int16_t * const B= h->mv_cache[list][ index8 - 8 ];
761     const int16_t * C;
762     int diagonal_ref, match_count;
763
764     assert(part_width==1 || part_width==2 || part_width==4);
765
766 /* mv_cache
767   B . . A T T T T
768   U . . L . . , .
769   U . . L . . . .
770   U . . L . . , .
771   . . . L . . . .
772 */
773
774     diagonal_ref= fetch_diagonal_mv(h, &C, index8, list, part_width);
775     match_count= (diagonal_ref==ref) + (top_ref==ref) + (left_ref==ref);
776     tprintf(h->s.avctx, "pred_motion match_count=%d\n", match_count);
777     if(match_count > 1){ //most common
778         *mx= mid_pred(A[0], B[0], C[0]);
779         *my= mid_pred(A[1], B[1], C[1]);
780     }else if(match_count==1){
781         if(left_ref==ref){
782             *mx= A[0];
783             *my= A[1];
784         }else if(top_ref==ref){
785             *mx= B[0];
786             *my= B[1];
787         }else{
788             *mx= C[0];
789             *my= C[1];
790         }
791     }else{
792         if(top_ref == PART_NOT_AVAILABLE && diagonal_ref == PART_NOT_AVAILABLE && left_ref != PART_NOT_AVAILABLE){
793             *mx= A[0];
794             *my= A[1];
795         }else{
796             *mx= mid_pred(A[0], B[0], C[0]);
797             *my= mid_pred(A[1], B[1], C[1]);
798         }
799     }
800
801     tprintf(h->s.avctx, "pred_motion (%2d %2d %2d) (%2d %2d %2d) (%2d %2d %2d) -> (%2d %2d %2d) at %2d %2d %d list %d\n", top_ref, B[0], B[1],                    diagonal_ref, C[0], C[1], left_ref, A[0], A[1], ref, *mx, *my, h->s.mb_x, h->s.mb_y, n, list);
802 }
803
804 /**
805  * gets the directionally predicted 16x8 MV.
806  * @param n the block index
807  * @param mx the x component of the predicted motion vector
808  * @param my the y component of the predicted motion vector
809  */
810 static inline void pred_16x8_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
811     if(n==0){
812         const int top_ref=      h->ref_cache[list][ scan8[0] - 8 ];
813         const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
814
815         tprintf(h->s.avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n", top_ref, B[0], B[1], h->s.mb_x, h->s.mb_y, n, list);
816
817         if(top_ref == ref){
818             *mx= B[0];
819             *my= B[1];
820             return;
821         }
822     }else{
823         const int left_ref=     h->ref_cache[list][ scan8[8] - 1 ];
824         const int16_t * const A= h->mv_cache[list][ scan8[8] - 1 ];
825
826         tprintf(h->s.avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n", left_ref, A[0], A[1], h->s.mb_x, h->s.mb_y, n, list);
827
828         if(left_ref == ref){
829             *mx= A[0];
830             *my= A[1];
831             return;
832         }
833     }
834
835     //RARE
836     pred_motion(h, n, 4, list, ref, mx, my);
837 }
838
839 /**
840  * gets the directionally predicted 8x16 MV.
841  * @param n the block index
842  * @param mx the x component of the predicted motion vector
843  * @param my the y component of the predicted motion vector
844  */
845 static inline void pred_8x16_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
846     if(n==0){
847         const int left_ref=      h->ref_cache[list][ scan8[0] - 1 ];
848         const int16_t * const A=  h->mv_cache[list][ scan8[0] - 1 ];
849
850         tprintf(h->s.avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n", left_ref, A[0], A[1], h->s.mb_x, h->s.mb_y, n, list);
851
852         if(left_ref == ref){
853             *mx= A[0];
854             *my= A[1];
855             return;
856         }
857     }else{
858         const int16_t * C;
859         int diagonal_ref;
860
861         diagonal_ref= fetch_diagonal_mv(h, &C, scan8[4], list, 2);
862
863         tprintf(h->s.avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n", diagonal_ref, C[0], C[1], h->s.mb_x, h->s.mb_y, n, list);
864
865         if(diagonal_ref == ref){
866             *mx= C[0];
867             *my= C[1];
868             return;
869         }
870     }
871
872     //RARE
873     pred_motion(h, n, 2, list, ref, mx, my);
874 }
875
876 static inline void pred_pskip_motion(H264Context * const h, int * const mx, int * const my){
877     const int top_ref = h->ref_cache[0][ scan8[0] - 8 ];
878     const int left_ref= h->ref_cache[0][ scan8[0] - 1 ];
879
880     tprintf(h->s.avctx, "pred_pskip: (%d) (%d) at %2d %2d\n", top_ref, left_ref, h->s.mb_x, h->s.mb_y);
881
882     if(top_ref == PART_NOT_AVAILABLE || left_ref == PART_NOT_AVAILABLE
883        || (top_ref == 0  && *(uint32_t*)h->mv_cache[0][ scan8[0] - 8 ] == 0)
884        || (left_ref == 0 && *(uint32_t*)h->mv_cache[0][ scan8[0] - 1 ] == 0)){
885
886         *mx = *my = 0;
887         return;
888     }
889
890     pred_motion(h, 0, 4, 0, 0, mx, my);
891
892     return;
893 }
894
895 static inline void direct_dist_scale_factor(H264Context * const h){
896     const int poc = h->s.current_picture_ptr->poc;
897     const int poc1 = h->ref_list[1][0].poc;
898     int i;
899     for(i=0; i<h->ref_count[0]; i++){
900         int poc0 = h->ref_list[0][i].poc;
901         int td = av_clip(poc1 - poc0, -128, 127);
902         if(td == 0 /* FIXME || pic0 is a long-term ref */){
903             h->dist_scale_factor[i] = 256;
904         }else{
905             int tb = av_clip(poc - poc0, -128, 127);
906             int tx = (16384 + (FFABS(td) >> 1)) / td;
907             h->dist_scale_factor[i] = av_clip((tb*tx + 32) >> 6, -1024, 1023);
908         }
909     }
910     if(FRAME_MBAFF){
911         for(i=0; i<h->ref_count[0]; i++){
912             h->dist_scale_factor_field[2*i] =
913             h->dist_scale_factor_field[2*i+1] = h->dist_scale_factor[i];
914         }
915     }
916 }
917 static inline void direct_ref_list_init(H264Context * const h){
918     MpegEncContext * const s = &h->s;
919     Picture * const ref1 = &h->ref_list[1][0];
920     Picture * const cur = s->current_picture_ptr;
921     int list, i, j;
922     if(cur->pict_type == FF_I_TYPE)
923         cur->ref_count[0] = 0;
924     if(cur->pict_type != FF_B_TYPE)
925         cur->ref_count[1] = 0;
926     for(list=0; list<2; list++){
927         cur->ref_count[list] = h->ref_count[list];
928         for(j=0; j<h->ref_count[list]; j++)
929             cur->ref_poc[list][j] = h->ref_list[list][j].poc;
930     }
931     if(cur->pict_type != FF_B_TYPE || h->direct_spatial_mv_pred)
932         return;
933     for(list=0; list<2; list++){
934         for(i=0; i<ref1->ref_count[list]; i++){
935             const int poc = ref1->ref_poc[list][i];
936             h->map_col_to_list0[list][i] = 0; /* bogus; fills in for missing frames */
937             for(j=0; j<h->ref_count[list]; j++)
938                 if(h->ref_list[list][j].poc == poc){
939                     h->map_col_to_list0[list][i] = j;
940                     break;
941                 }
942         }
943     }
944     if(FRAME_MBAFF){
945         for(list=0; list<2; list++){
946             for(i=0; i<ref1->ref_count[list]; i++){
947                 j = h->map_col_to_list0[list][i];
948                 h->map_col_to_list0_field[list][2*i] = 2*j;
949                 h->map_col_to_list0_field[list][2*i+1] = 2*j+1;
950             }
951         }
952     }
953 }
954
955 static inline void pred_direct_motion(H264Context * const h, int *mb_type){
956     MpegEncContext * const s = &h->s;
957     const int mb_xy =   s->mb_x +   s->mb_y*s->mb_stride;
958     const int b8_xy = 2*s->mb_x + 2*s->mb_y*h->b8_stride;
959     const int b4_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
960     const int mb_type_col = h->ref_list[1][0].mb_type[mb_xy];
961     const int16_t (*l1mv0)[2] = (const int16_t (*)[2]) &h->ref_list[1][0].motion_val[0][b4_xy];
962     const int16_t (*l1mv1)[2] = (const int16_t (*)[2]) &h->ref_list[1][0].motion_val[1][b4_xy];
963     const int8_t *l1ref0 = &h->ref_list[1][0].ref_index[0][b8_xy];
964     const int8_t *l1ref1 = &h->ref_list[1][0].ref_index[1][b8_xy];
965     const int is_b8x8 = IS_8X8(*mb_type);
966     unsigned int sub_mb_type;
967     int i8, i4;
968
969 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
970     if(IS_8X8(mb_type_col) && !h->sps.direct_8x8_inference_flag){
971         /* FIXME save sub mb types from previous frames (or derive from MVs)
972          * so we know exactly what block size to use */
973         sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
974         *mb_type =    MB_TYPE_8x8|MB_TYPE_L0L1;
975     }else if(!is_b8x8 && (mb_type_col & MB_TYPE_16x16_OR_INTRA)){
976         sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
977         *mb_type =    MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
978     }else{
979         sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
980         *mb_type =    MB_TYPE_8x8|MB_TYPE_L0L1;
981     }
982     if(!is_b8x8)
983         *mb_type |= MB_TYPE_DIRECT2;
984     if(MB_FIELD)
985         *mb_type |= MB_TYPE_INTERLACED;
986
987     tprintf(s->avctx, "mb_type = %08x, sub_mb_type = %08x, is_b8x8 = %d, mb_type_col = %08x\n", *mb_type, sub_mb_type, is_b8x8, mb_type_col);
988
989     if(h->direct_spatial_mv_pred){
990         int ref[2];
991         int mv[2][2];
992         int list;
993
994         /* FIXME interlacing + spatial direct uses wrong colocated block positions */
995
996         /* ref = min(neighbors) */
997         for(list=0; list<2; list++){
998             int refa = h->ref_cache[list][scan8[0] - 1];
999             int refb = h->ref_cache[list][scan8[0] - 8];
1000             int refc = h->ref_cache[list][scan8[0] - 8 + 4];
1001             if(refc == -2)
1002                 refc = h->ref_cache[list][scan8[0] - 8 - 1];
1003             ref[list] = refa;
1004             if(ref[list] < 0 || (refb < ref[list] && refb >= 0))
1005                 ref[list] = refb;
1006             if(ref[list] < 0 || (refc < ref[list] && refc >= 0))
1007                 ref[list] = refc;
1008             if(ref[list] < 0)
1009                 ref[list] = -1;
1010         }
1011
1012         if(ref[0] < 0 && ref[1] < 0){
1013             ref[0] = ref[1] = 0;
1014             mv[0][0] = mv[0][1] =
1015             mv[1][0] = mv[1][1] = 0;
1016         }else{
1017             for(list=0; list<2; list++){
1018                 if(ref[list] >= 0)
1019                     pred_motion(h, 0, 4, list, ref[list], &mv[list][0], &mv[list][1]);
1020                 else
1021                     mv[list][0] = mv[list][1] = 0;
1022             }
1023         }
1024
1025         if(ref[1] < 0){
1026             if(!is_b8x8)
1027                 *mb_type &= ~MB_TYPE_L1;
1028             sub_mb_type &= ~MB_TYPE_L1;
1029         }else if(ref[0] < 0){
1030             if(!is_b8x8)
1031                 *mb_type &= ~MB_TYPE_L0;
1032             sub_mb_type &= ~MB_TYPE_L0;
1033         }
1034
1035         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col)){
1036             int pair_xy = s->mb_x + (s->mb_y&~1)*s->mb_stride;
1037             int mb_types_col[2];
1038             int b8_stride = h->b8_stride;
1039             int b4_stride = h->b_stride;
1040
1041             *mb_type = (*mb_type & ~MB_TYPE_16x16) | MB_TYPE_8x8;
1042
1043             if(IS_INTERLACED(*mb_type)){
1044                 mb_types_col[0] = h->ref_list[1][0].mb_type[pair_xy];
1045                 mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1046                 if(s->mb_y&1){
1047                     l1ref0 -= 2*b8_stride;
1048                     l1ref1 -= 2*b8_stride;
1049                     l1mv0 -= 4*b4_stride;
1050                     l1mv1 -= 4*b4_stride;
1051                 }
1052                 b8_stride *= 3;
1053                 b4_stride *= 6;
1054             }else{
1055                 int cur_poc = s->current_picture_ptr->poc;
1056                 int *col_poc = h->ref_list[1]->field_poc;
1057                 int col_parity = FFABS(col_poc[0] - cur_poc) >= FFABS(col_poc[1] - cur_poc);
1058                 int dy = 2*col_parity - (s->mb_y&1);
1059                 mb_types_col[0] =
1060                 mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy + col_parity*s->mb_stride];
1061                 l1ref0 += dy*b8_stride;
1062                 l1ref1 += dy*b8_stride;
1063                 l1mv0 += 2*dy*b4_stride;
1064                 l1mv1 += 2*dy*b4_stride;
1065                 b8_stride = 0;
1066             }
1067
1068             for(i8=0; i8<4; i8++){
1069                 int x8 = i8&1;
1070                 int y8 = i8>>1;
1071                 int xy8 = x8+y8*b8_stride;
1072                 int xy4 = 3*x8+y8*b4_stride;
1073                 int a=0, b=0;
1074
1075                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1076                     continue;
1077                 h->sub_mb_type[i8] = sub_mb_type;
1078
1079                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1080                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1081                 if(!IS_INTRA(mb_types_col[y8])
1082                    && (   (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
1083                        || (l1ref0[xy8]  < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
1084                     if(ref[0] > 0)
1085                         a= pack16to32(mv[0][0],mv[0][1]);
1086                     if(ref[1] > 0)
1087                         b= pack16to32(mv[1][0],mv[1][1]);
1088                 }else{
1089                     a= pack16to32(mv[0][0],mv[0][1]);
1090                     b= pack16to32(mv[1][0],mv[1][1]);
1091                 }
1092                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
1093                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
1094             }
1095         }else if(IS_16X16(*mb_type)){
1096             int a=0, b=0;
1097
1098             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
1099             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
1100             if(!IS_INTRA(mb_type_col)
1101                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
1102                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
1103                        && (h->x264_build>33 || !h->x264_build)))){
1104                 if(ref[0] > 0)
1105                     a= pack16to32(mv[0][0],mv[0][1]);
1106                 if(ref[1] > 0)
1107                     b= pack16to32(mv[1][0],mv[1][1]);
1108             }else{
1109                 a= pack16to32(mv[0][0],mv[0][1]);
1110                 b= pack16to32(mv[1][0],mv[1][1]);
1111             }
1112             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
1113             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
1114         }else{
1115             for(i8=0; i8<4; i8++){
1116                 const int x8 = i8&1;
1117                 const int y8 = i8>>1;
1118
1119                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1120                     continue;
1121                 h->sub_mb_type[i8] = sub_mb_type;
1122
1123                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mv[0][0],mv[0][1]), 4);
1124                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mv[1][0],mv[1][1]), 4);
1125                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1126                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1127
1128                 /* col_zero_flag */
1129                 if(!IS_INTRA(mb_type_col) && (   l1ref0[x8 + y8*h->b8_stride] == 0
1130                                               || (l1ref0[x8 + y8*h->b8_stride] < 0 && l1ref1[x8 + y8*h->b8_stride] == 0
1131                                                   && (h->x264_build>33 || !h->x264_build)))){
1132                     const int16_t (*l1mv)[2]= l1ref0[x8 + y8*h->b8_stride] == 0 ? l1mv0 : l1mv1;
1133                     if(IS_SUB_8X8(sub_mb_type)){
1134                         const int16_t *mv_col = l1mv[x8*3 + y8*3*h->b_stride];
1135                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1136                             if(ref[0] == 0)
1137                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1138                             if(ref[1] == 0)
1139                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1140                         }
1141                     }else
1142                     for(i4=0; i4<4; i4++){
1143                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*h->b_stride];
1144                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1145                             if(ref[0] == 0)
1146                                 *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
1147                             if(ref[1] == 0)
1148                                 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
1149                         }
1150                     }
1151                 }
1152             }
1153         }
1154     }else{ /* direct temporal mv pred */
1155         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
1156         const int *dist_scale_factor = h->dist_scale_factor;
1157
1158         if(FRAME_MBAFF){
1159             if(IS_INTERLACED(*mb_type)){
1160                 map_col_to_list0[0] = h->map_col_to_list0_field[0];
1161                 map_col_to_list0[1] = h->map_col_to_list0_field[1];
1162                 dist_scale_factor = h->dist_scale_factor_field;
1163             }
1164             if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col)){
1165                 /* FIXME assumes direct_8x8_inference == 1 */
1166                 const int pair_xy = s->mb_x + (s->mb_y&~1)*s->mb_stride;
1167                 int mb_types_col[2];
1168                 int y_shift;
1169
1170                 *mb_type = MB_TYPE_8x8|MB_TYPE_L0L1
1171                          | (is_b8x8 ? 0 : MB_TYPE_DIRECT2)
1172                          | (*mb_type & MB_TYPE_INTERLACED);
1173                 sub_mb_type = MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2|MB_TYPE_16x16;
1174
1175                 if(IS_INTERLACED(*mb_type)){
1176                     /* frame to field scaling */
1177                     mb_types_col[0] = h->ref_list[1][0].mb_type[pair_xy];
1178                     mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1179                     if(s->mb_y&1){
1180                         l1ref0 -= 2*h->b8_stride;
1181                         l1ref1 -= 2*h->b8_stride;
1182                         l1mv0 -= 4*h->b_stride;
1183                         l1mv1 -= 4*h->b_stride;
1184                     }
1185                     y_shift = 0;
1186
1187                     if(   (mb_types_col[0] & MB_TYPE_16x16_OR_INTRA)
1188                        && (mb_types_col[1] & MB_TYPE_16x16_OR_INTRA)
1189                        && !is_b8x8)
1190                         *mb_type |= MB_TYPE_16x8;
1191                     else
1192                         *mb_type |= MB_TYPE_8x8;
1193                 }else{
1194                     /* field to frame scaling */
1195                     /* col_mb_y = (mb_y&~1) + (topAbsDiffPOC < bottomAbsDiffPOC ? 0 : 1)
1196                      * but in MBAFF, top and bottom POC are equal */
1197                     int dy = (s->mb_y&1) ? 1 : 2;
1198                     mb_types_col[0] =
1199                     mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1200                     l1ref0 += dy*h->b8_stride;
1201                     l1ref1 += dy*h->b8_stride;
1202                     l1mv0 += 2*dy*h->b_stride;
1203                     l1mv1 += 2*dy*h->b_stride;
1204                     y_shift = 2;
1205
1206                     if((mb_types_col[0] & (MB_TYPE_16x16_OR_INTRA|MB_TYPE_16x8))
1207                        && !is_b8x8)
1208                         *mb_type |= MB_TYPE_16x16;
1209                     else
1210                         *mb_type |= MB_TYPE_8x8;
1211                 }
1212
1213                 for(i8=0; i8<4; i8++){
1214                     const int x8 = i8&1;
1215                     const int y8 = i8>>1;
1216                     int ref0, scale;
1217                     const int16_t (*l1mv)[2]= l1mv0;
1218
1219                     if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1220                         continue;
1221                     h->sub_mb_type[i8] = sub_mb_type;
1222
1223                     fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1224                     if(IS_INTRA(mb_types_col[y8])){
1225                         fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1226                         fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1227                         fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1228                         continue;
1229                     }
1230
1231                     ref0 = l1ref0[x8 + (y8*2>>y_shift)*h->b8_stride];
1232                     if(ref0 >= 0)
1233                         ref0 = map_col_to_list0[0][ref0*2>>y_shift];
1234                     else{
1235                         ref0 = map_col_to_list0[1][l1ref1[x8 + (y8*2>>y_shift)*h->b8_stride]*2>>y_shift];
1236                         l1mv= l1mv1;
1237                     }
1238                     scale = dist_scale_factor[ref0];
1239                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1240
1241                     {
1242                         const int16_t *mv_col = l1mv[x8*3 + (y8*6>>y_shift)*h->b_stride];
1243                         int my_col = (mv_col[1]<<y_shift)/2;
1244                         int mx = (scale * mv_col[0] + 128) >> 8;
1245                         int my = (scale * my_col + 128) >> 8;
1246                         fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1247                         fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
1248                     }
1249                 }
1250                 return;
1251             }
1252         }
1253
1254         /* one-to-one mv scaling */
1255
1256         if(IS_16X16(*mb_type)){
1257             int ref, mv0, mv1;
1258
1259             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
1260             if(IS_INTRA(mb_type_col)){
1261                 ref=mv0=mv1=0;
1262             }else{
1263                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0]]
1264                                                 : map_col_to_list0[1][l1ref1[0]];
1265                 const int scale = dist_scale_factor[ref0];
1266                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
1267                 int mv_l0[2];
1268                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1269                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1270                 ref= ref0;
1271                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
1272                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1273             }
1274             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
1275             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
1276             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
1277         }else{
1278             for(i8=0; i8<4; i8++){
1279                 const int x8 = i8&1;
1280                 const int y8 = i8>>1;
1281                 int ref0, scale;
1282                 const int16_t (*l1mv)[2]= l1mv0;
1283
1284                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1285                     continue;
1286                 h->sub_mb_type[i8] = sub_mb_type;
1287                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1288                 if(IS_INTRA(mb_type_col)){
1289                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1290                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1291                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1292                     continue;
1293                 }
1294
1295                 ref0 = l1ref0[x8 + y8*h->b8_stride];
1296                 if(ref0 >= 0)
1297                     ref0 = map_col_to_list0[0][ref0];
1298                 else{
1299                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*h->b8_stride]];
1300                     l1mv= l1mv1;
1301                 }
1302                 scale = dist_scale_factor[ref0];
1303
1304                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1305                 if(IS_SUB_8X8(sub_mb_type)){
1306                     const int16_t *mv_col = l1mv[x8*3 + y8*3*h->b_stride];
1307                     int mx = (scale * mv_col[0] + 128) >> 8;
1308                     int my = (scale * mv_col[1] + 128) >> 8;
1309                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1310                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
1311                 }else
1312                 for(i4=0; i4<4; i4++){
1313                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*h->b_stride];
1314                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
1315                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1316                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1317                     *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
1318                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1319                 }
1320             }
1321         }
1322     }
1323 }
1324
1325 static inline void write_back_motion(H264Context *h, int mb_type){
1326     MpegEncContext * const s = &h->s;
1327     const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1328     const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1329     int list;
1330
1331     if(!USES_LIST(mb_type, 0))
1332         fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
1333
1334     for(list=0; list<h->list_count; list++){
1335         int y;
1336         if(!USES_LIST(mb_type, list))
1337             continue;
1338
1339         for(y=0; y<4; y++){
1340             *(uint64_t*)s->current_picture.motion_val[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mv_cache[list][scan8[0]+0 + 8*y];
1341             *(uint64_t*)s->current_picture.motion_val[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mv_cache[list][scan8[0]+2 + 8*y];
1342         }
1343         if( h->pps.cabac ) {
1344             if(IS_SKIP(mb_type))
1345                 fill_rectangle(h->mvd_table[list][b_xy], 4, 4, h->b_stride, 0, 4);
1346             else
1347             for(y=0; y<4; y++){
1348                 *(uint64_t*)h->mvd_table[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+0 + 8*y];
1349                 *(uint64_t*)h->mvd_table[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+2 + 8*y];
1350             }
1351         }
1352
1353         {
1354             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
1355             ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
1356             ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
1357             ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
1358             ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
1359         }
1360     }
1361
1362     if(h->slice_type == FF_B_TYPE && h->pps.cabac){
1363         if(IS_8X8(mb_type)){
1364             uint8_t *direct_table = &h->direct_table[b8_xy];
1365             direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
1366             direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
1367             direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
1368         }
1369     }
1370 }
1371
1372 /**
1373  * Decodes a network abstraction layer unit.
1374  * @param consumed is the number of bytes used as input
1375  * @param length is the length of the array
1376  * @param dst_length is the number of decoded bytes FIXME here or a decode rbsp tailing?
1377  * @returns decoded bytes, might be src+1 if no escapes
1378  */
1379 static const uint8_t *decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
1380     int i, si, di;
1381     uint8_t *dst;
1382     int bufidx;
1383
1384 //    src[0]&0x80;                //forbidden bit
1385     h->nal_ref_idc= src[0]>>5;
1386     h->nal_unit_type= src[0]&0x1F;
1387
1388     src++; length--;
1389 #if 0
1390     for(i=0; i<length; i++)
1391         printf("%2X ", src[i]);
1392 #endif
1393     for(i=0; i+1<length; i+=2){
1394         if(src[i]) continue;
1395         if(i>0 && src[i-1]==0) i--;
1396         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
1397             if(src[i+2]!=3){
1398                 /* startcode, so we must be past the end */
1399                 length=i;
1400             }
1401             break;
1402         }
1403     }
1404
1405     if(i>=length-1){ //no escaped 0
1406         *dst_length= length;
1407         *consumed= length+1; //+1 for the header
1408         return src;
1409     }
1410
1411     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
1412     h->rbsp_buffer[bufidx]= av_fast_realloc(h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length);
1413     dst= h->rbsp_buffer[bufidx];
1414
1415     if (dst == NULL){
1416         return NULL;
1417     }
1418
1419 //printf("decoding esc\n");
1420     si=di=0;
1421     while(si<length){
1422         //remove escapes (very rare 1:2^22)
1423         if(si+2<length && src[si]==0 && src[si+1]==0 && src[si+2]<=3){
1424             if(src[si+2]==3){ //escape
1425                 dst[di++]= 0;
1426                 dst[di++]= 0;
1427                 si+=3;
1428                 continue;
1429             }else //next start code
1430                 break;
1431         }
1432
1433         dst[di++]= src[si++];
1434     }
1435
1436     *dst_length= di;
1437     *consumed= si + 1;//+1 for the header
1438 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
1439     return dst;
1440 }
1441
1442 /**
1443  * identifies the exact end of the bitstream
1444  * @return the length of the trailing, or 0 if damaged
1445  */
1446 static int decode_rbsp_trailing(H264Context *h, const uint8_t *src){
1447     int v= *src;
1448     int r;
1449
1450     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
1451
1452     for(r=1; r<9; r++){
1453         if(v&1) return r;
1454         v>>=1;
1455     }
1456     return 0;
1457 }
1458
1459 /**
1460  * idct tranforms the 16 dc values and dequantize them.
1461  * @param qp quantization parameter
1462  */
1463 static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1464 #define stride 16
1465     int i;
1466     int temp[16]; //FIXME check if this is a good idea
1467     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1468     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1469
1470 //memset(block, 64, 2*256);
1471 //return;
1472     for(i=0; i<4; i++){
1473         const int offset= y_offset[i];
1474         const int z0= block[offset+stride*0] + block[offset+stride*4];
1475         const int z1= block[offset+stride*0] - block[offset+stride*4];
1476         const int z2= block[offset+stride*1] - block[offset+stride*5];
1477         const int z3= block[offset+stride*1] + block[offset+stride*5];
1478
1479         temp[4*i+0]= z0+z3;
1480         temp[4*i+1]= z1+z2;
1481         temp[4*i+2]= z1-z2;
1482         temp[4*i+3]= z0-z3;
1483     }
1484
1485     for(i=0; i<4; i++){
1486         const int offset= x_offset[i];
1487         const int z0= temp[4*0+i] + temp[4*2+i];
1488         const int z1= temp[4*0+i] - temp[4*2+i];
1489         const int z2= temp[4*1+i] - temp[4*3+i];
1490         const int z3= temp[4*1+i] + temp[4*3+i];
1491
1492         block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_resdual
1493         block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
1494         block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
1495         block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
1496     }
1497 }
1498
1499 #if 0
1500 /**
1501  * dct tranforms the 16 dc values.
1502  * @param qp quantization parameter ??? FIXME
1503  */
1504 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
1505 //    const int qmul= dequant_coeff[qp][0];
1506     int i;
1507     int temp[16]; //FIXME check if this is a good idea
1508     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1509     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1510
1511     for(i=0; i<4; i++){
1512         const int offset= y_offset[i];
1513         const int z0= block[offset+stride*0] + block[offset+stride*4];
1514         const int z1= block[offset+stride*0] - block[offset+stride*4];
1515         const int z2= block[offset+stride*1] - block[offset+stride*5];
1516         const int z3= block[offset+stride*1] + block[offset+stride*5];
1517
1518         temp[4*i+0]= z0+z3;
1519         temp[4*i+1]= z1+z2;
1520         temp[4*i+2]= z1-z2;
1521         temp[4*i+3]= z0-z3;
1522     }
1523
1524     for(i=0; i<4; i++){
1525         const int offset= x_offset[i];
1526         const int z0= temp[4*0+i] + temp[4*2+i];
1527         const int z1= temp[4*0+i] - temp[4*2+i];
1528         const int z2= temp[4*1+i] - temp[4*3+i];
1529         const int z3= temp[4*1+i] + temp[4*3+i];
1530
1531         block[stride*0 +offset]= (z0 + z3)>>1;
1532         block[stride*2 +offset]= (z1 + z2)>>1;
1533         block[stride*8 +offset]= (z1 - z2)>>1;
1534         block[stride*10+offset]= (z0 - z3)>>1;
1535     }
1536 }
1537 #endif
1538
1539 #undef xStride
1540 #undef stride
1541
1542 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1543     const int stride= 16*2;
1544     const int xStride= 16;
1545     int a,b,c,d,e;
1546
1547     a= block[stride*0 + xStride*0];
1548     b= block[stride*0 + xStride*1];
1549     c= block[stride*1 + xStride*0];
1550     d= block[stride*1 + xStride*1];
1551
1552     e= a-b;
1553     a= a+b;
1554     b= c-d;
1555     c= c+d;
1556
1557     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
1558     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
1559     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
1560     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
1561 }
1562
1563 #if 0
1564 static void chroma_dc_dct_c(DCTELEM *block){
1565     const int stride= 16*2;
1566     const int xStride= 16;
1567     int a,b,c,d,e;
1568
1569     a= block[stride*0 + xStride*0];
1570     b= block[stride*0 + xStride*1];
1571     c= block[stride*1 + xStride*0];
1572     d= block[stride*1 + xStride*1];
1573
1574     e= a-b;
1575     a= a+b;
1576     b= c-d;
1577     c= c+d;
1578
1579     block[stride*0 + xStride*0]= (a+c);
1580     block[stride*0 + xStride*1]= (e+b);
1581     block[stride*1 + xStride*0]= (a-c);
1582     block[stride*1 + xStride*1]= (e-b);
1583 }
1584 #endif
1585
1586 /**
1587  * gets the chroma qp.
1588  */
1589 static inline int get_chroma_qp(H264Context *h, int t, int qscale){
1590     return h->pps.chroma_qp_table[t][qscale & 0xff];
1591 }
1592
1593 //FIXME need to check that this does not overflow signed 32 bit for low qp, I am not sure, it's very close
1594 //FIXME check that gcc inlines this (and optimizes intra & separate_dc stuff away)
1595 static inline int quantize_c(DCTELEM *block, uint8_t *scantable, int qscale, int intra, int separate_dc){
1596     int i;
1597     const int * const quant_table= quant_coeff[qscale];
1598     const int bias= intra ? (1<<QUANT_SHIFT)/3 : (1<<QUANT_SHIFT)/6;
1599     const unsigned int threshold1= (1<<QUANT_SHIFT) - bias - 1;
1600     const unsigned int threshold2= (threshold1<<1);
1601     int last_non_zero;
1602
1603     if(separate_dc){
1604         if(qscale<=18){
1605             //avoid overflows
1606             const int dc_bias= intra ? (1<<(QUANT_SHIFT-2))/3 : (1<<(QUANT_SHIFT-2))/6;
1607             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT-2)) - dc_bias - 1;
1608             const unsigned int dc_threshold2= (dc_threshold1<<1);
1609
1610             int level= block[0]*quant_coeff[qscale+18][0];
1611             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1612                 if(level>0){
1613                     level= (dc_bias + level)>>(QUANT_SHIFT-2);
1614                     block[0]= level;
1615                 }else{
1616                     level= (dc_bias - level)>>(QUANT_SHIFT-2);
1617                     block[0]= -level;
1618                 }
1619 //                last_non_zero = i;
1620             }else{
1621                 block[0]=0;
1622             }
1623         }else{
1624             const int dc_bias= intra ? (1<<(QUANT_SHIFT+1))/3 : (1<<(QUANT_SHIFT+1))/6;
1625             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT+1)) - dc_bias - 1;
1626             const unsigned int dc_threshold2= (dc_threshold1<<1);
1627
1628             int level= block[0]*quant_table[0];
1629             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1630                 if(level>0){
1631                     level= (dc_bias + level)>>(QUANT_SHIFT+1);
1632                     block[0]= level;
1633                 }else{
1634                     level= (dc_bias - level)>>(QUANT_SHIFT+1);
1635                     block[0]= -level;
1636                 }
1637 //                last_non_zero = i;
1638             }else{
1639                 block[0]=0;
1640             }
1641         }
1642         last_non_zero= 0;
1643         i=1;
1644     }else{
1645         last_non_zero= -1;
1646         i=0;
1647     }
1648
1649     for(; i<16; i++){
1650         const int j= scantable[i];
1651         int level= block[j]*quant_table[j];
1652
1653 //        if(   bias+level >= (1<<(QMAT_SHIFT - 3))
1654 //           || bias-level >= (1<<(QMAT_SHIFT - 3))){
1655         if(((unsigned)(level+threshold1))>threshold2){
1656             if(level>0){
1657                 level= (bias + level)>>QUANT_SHIFT;
1658                 block[j]= level;
1659             }else{
1660                 level= (bias - level)>>QUANT_SHIFT;
1661                 block[j]= -level;
1662             }
1663             last_non_zero = i;
1664         }else{
1665             block[j]=0;
1666         }
1667     }
1668
1669     return last_non_zero;
1670 }
1671
1672 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
1673                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1674                            int src_x_offset, int src_y_offset,
1675                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
1676     MpegEncContext * const s = &h->s;
1677     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
1678     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
1679     const int luma_xy= (mx&3) + ((my&3)<<2);
1680     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
1681     uint8_t * src_cb, * src_cr;
1682     int extra_width= h->emu_edge_width;
1683     int extra_height= h->emu_edge_height;
1684     int emu=0;
1685     const int full_mx= mx>>2;
1686     const int full_my= my>>2;
1687     const int pic_width  = 16*s->mb_width;
1688     const int pic_height = 16*s->mb_height >> MB_FIELD;
1689
1690     if(!pic->data[0]) //FIXME this is unacceptable, some senseable error concealment must be done for missing reference frames
1691         return;
1692
1693     if(mx&7) extra_width -= 3;
1694     if(my&7) extra_height -= 3;
1695
1696     if(   full_mx < 0-extra_width
1697        || full_my < 0-extra_height
1698        || full_mx + 16/*FIXME*/ > pic_width + extra_width
1699        || full_my + 16/*FIXME*/ > pic_height + extra_height){
1700         ff_emulated_edge_mc(s->edge_emu_buffer, src_y - 2 - 2*h->mb_linesize, h->mb_linesize, 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
1701             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
1702         emu=1;
1703     }
1704
1705     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
1706     if(!square){
1707         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
1708     }
1709
1710     if(ENABLE_GRAY && s->flags&CODEC_FLAG_GRAY) return;
1711
1712     if(MB_FIELD){
1713         // chroma offset when predicting from a field of opposite parity
1714         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
1715         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
1716     }
1717     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1718     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1719
1720     if(emu){
1721         ff_emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
1722             src_cb= s->edge_emu_buffer;
1723     }
1724     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1725
1726     if(emu){
1727         ff_emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
1728             src_cr= s->edge_emu_buffer;
1729     }
1730     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1731 }
1732
1733 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
1734                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1735                            int x_offset, int y_offset,
1736                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1737                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1738                            int list0, int list1){
1739     MpegEncContext * const s = &h->s;
1740     qpel_mc_func *qpix_op=  qpix_put;
1741     h264_chroma_mc_func chroma_op= chroma_put;
1742
1743     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1744     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1745     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1746     x_offset += 8*s->mb_x;
1747     y_offset += 8*(s->mb_y >> MB_FIELD);
1748
1749     if(list0){
1750         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
1751         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
1752                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1753                            qpix_op, chroma_op);
1754
1755         qpix_op=  qpix_avg;
1756         chroma_op= chroma_avg;
1757     }
1758
1759     if(list1){
1760         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
1761         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
1762                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1763                            qpix_op, chroma_op);
1764     }
1765 }
1766
1767 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
1768                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1769                            int x_offset, int y_offset,
1770                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1771                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
1772                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
1773                            int list0, int list1){
1774     MpegEncContext * const s = &h->s;
1775
1776     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1777     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1778     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1779     x_offset += 8*s->mb_x;
1780     y_offset += 8*(s->mb_y >> MB_FIELD);
1781
1782     if(list0 && list1){
1783         /* don't optimize for luma-only case, since B-frames usually
1784          * use implicit weights => chroma too. */
1785         uint8_t *tmp_cb = s->obmc_scratchpad;
1786         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
1787         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
1788         int refn0 = h->ref_cache[0][ scan8[n] ];
1789         int refn1 = h->ref_cache[1][ scan8[n] ];
1790
1791         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
1792                     dest_y, dest_cb, dest_cr,
1793                     x_offset, y_offset, qpix_put, chroma_put);
1794         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
1795                     tmp_y, tmp_cb, tmp_cr,
1796                     x_offset, y_offset, qpix_put, chroma_put);
1797
1798         if(h->use_weight == 2){
1799             int weight0 = h->implicit_weight[refn0][refn1];
1800             int weight1 = 64 - weight0;
1801             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
1802             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
1803             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
1804         }else{
1805             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
1806                             h->luma_weight[0][refn0], h->luma_weight[1][refn1],
1807                             h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
1808             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1809                             h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
1810                             h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
1811             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1812                             h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
1813                             h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
1814         }
1815     }else{
1816         int list = list1 ? 1 : 0;
1817         int refn = h->ref_cache[list][ scan8[n] ];
1818         Picture *ref= &h->ref_list[list][refn];
1819         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
1820                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
1821                     qpix_put, chroma_put);
1822
1823         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
1824                        h->luma_weight[list][refn], h->luma_offset[list][refn]);
1825         if(h->use_weight_chroma){
1826             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1827                              h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
1828             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1829                              h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
1830         }
1831     }
1832 }
1833
1834 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
1835                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1836                            int x_offset, int y_offset,
1837                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1838                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1839                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
1840                            int list0, int list1){
1841     if((h->use_weight==2 && list0 && list1
1842         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
1843        || h->use_weight==1)
1844         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1845                          x_offset, y_offset, qpix_put, chroma_put,
1846                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
1847     else
1848         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1849                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
1850 }
1851
1852 static inline void prefetch_motion(H264Context *h, int list){
1853     /* fetch pixels for estimated mv 4 macroblocks ahead
1854      * optimized for 64byte cache lines */
1855     MpegEncContext * const s = &h->s;
1856     const int refn = h->ref_cache[list][scan8[0]];
1857     if(refn >= 0){
1858         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
1859         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
1860         uint8_t **src= h->ref_list[list][refn].data;
1861         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
1862         s->dsp.prefetch(src[0]+off, s->linesize, 4);
1863         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
1864         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1865     }
1866 }
1867
1868 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1869                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
1870                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
1871                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
1872     MpegEncContext * const s = &h->s;
1873     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
1874     const int mb_type= s->current_picture.mb_type[mb_xy];
1875
1876     assert(IS_INTER(mb_type));
1877
1878     prefetch_motion(h, 0);
1879
1880     if(IS_16X16(mb_type)){
1881         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
1882                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
1883                 &weight_op[0], &weight_avg[0],
1884                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1885     }else if(IS_16X8(mb_type)){
1886         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
1887                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1888                 &weight_op[1], &weight_avg[1],
1889                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1890         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
1891                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1892                 &weight_op[1], &weight_avg[1],
1893                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1894     }else if(IS_8X16(mb_type)){
1895         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
1896                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1897                 &weight_op[2], &weight_avg[2],
1898                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1899         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
1900                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1901                 &weight_op[2], &weight_avg[2],
1902                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1903     }else{
1904         int i;
1905
1906         assert(IS_8X8(mb_type));
1907
1908         for(i=0; i<4; i++){
1909             const int sub_mb_type= h->sub_mb_type[i];
1910             const int n= 4*i;
1911             int x_offset= (i&1)<<2;
1912             int y_offset= (i&2)<<1;
1913
1914             if(IS_SUB_8X8(sub_mb_type)){
1915                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1916                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1917                     &weight_op[3], &weight_avg[3],
1918                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1919             }else if(IS_SUB_8X4(sub_mb_type)){
1920                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1921                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1922                     &weight_op[4], &weight_avg[4],
1923                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1924                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
1925                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1926                     &weight_op[4], &weight_avg[4],
1927                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1928             }else if(IS_SUB_4X8(sub_mb_type)){
1929                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1930                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1931                     &weight_op[5], &weight_avg[5],
1932                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1933                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
1934                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1935                     &weight_op[5], &weight_avg[5],
1936                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1937             }else{
1938                 int j;
1939                 assert(IS_SUB_4X4(sub_mb_type));
1940                 for(j=0; j<4; j++){
1941                     int sub_x_offset= x_offset + 2*(j&1);
1942                     int sub_y_offset= y_offset +   (j&2);
1943                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
1944                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1945                         &weight_op[6], &weight_avg[6],
1946                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1947                 }
1948             }
1949         }
1950     }
1951
1952     prefetch_motion(h, 1);
1953 }
1954
1955 static void decode_init_vlc(void){
1956     static int done = 0;
1957
1958     if (!done) {
1959         int i;
1960         done = 1;
1961
1962         init_vlc(&chroma_dc_coeff_token_vlc, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 4*5,
1963                  &chroma_dc_coeff_token_len [0], 1, 1,
1964                  &chroma_dc_coeff_token_bits[0], 1, 1, 1);
1965
1966         for(i=0; i<4; i++){
1967             init_vlc(&coeff_token_vlc[i], COEFF_TOKEN_VLC_BITS, 4*17,
1968                      &coeff_token_len [i][0], 1, 1,
1969                      &coeff_token_bits[i][0], 1, 1, 1);
1970         }
1971
1972         for(i=0; i<3; i++){
1973             init_vlc(&chroma_dc_total_zeros_vlc[i], CHROMA_DC_TOTAL_ZEROS_VLC_BITS, 4,
1974                      &chroma_dc_total_zeros_len [i][0], 1, 1,
1975                      &chroma_dc_total_zeros_bits[i][0], 1, 1, 1);
1976         }
1977         for(i=0; i<15; i++){
1978             init_vlc(&total_zeros_vlc[i], TOTAL_ZEROS_VLC_BITS, 16,
1979                      &total_zeros_len [i][0], 1, 1,
1980                      &total_zeros_bits[i][0], 1, 1, 1);
1981         }
1982
1983         for(i=0; i<6; i++){
1984             init_vlc(&run_vlc[i], RUN_VLC_BITS, 7,
1985                      &run_len [i][0], 1, 1,
1986                      &run_bits[i][0], 1, 1, 1);
1987         }
1988         init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
1989                  &run_len [6][0], 1, 1,
1990                  &run_bits[6][0], 1, 1, 1);
1991     }
1992 }
1993
1994 static void free_tables(H264Context *h){
1995     int i;
1996     H264Context *hx;
1997     av_freep(&h->intra4x4_pred_mode);
1998     av_freep(&h->chroma_pred_mode_table);
1999     av_freep(&h->cbp_table);
2000     av_freep(&h->mvd_table[0]);
2001     av_freep(&h->mvd_table[1]);
2002     av_freep(&h->direct_table);
2003     av_freep(&h->non_zero_count);
2004     av_freep(&h->slice_table_base);
2005     h->slice_table= NULL;
2006
2007     av_freep(&h->mb2b_xy);
2008     av_freep(&h->mb2b8_xy);
2009
2010     for(i = 0; i < MAX_SPS_COUNT; i++)
2011         av_freep(h->sps_buffers + i);
2012
2013     for(i = 0; i < MAX_PPS_COUNT; i++)
2014         av_freep(h->pps_buffers + i);
2015
2016     for(i = 0; i < h->s.avctx->thread_count; i++) {
2017         hx = h->thread_context[i];
2018         if(!hx) continue;
2019         av_freep(&hx->top_borders[1]);
2020         av_freep(&hx->top_borders[0]);
2021         av_freep(&hx->s.obmc_scratchpad);
2022     }
2023 }
2024
2025 static void init_dequant8_coeff_table(H264Context *h){
2026     int i,q,x;
2027     const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
2028     h->dequant8_coeff[0] = h->dequant8_buffer[0];
2029     h->dequant8_coeff[1] = h->dequant8_buffer[1];
2030
2031     for(i=0; i<2; i++ ){
2032         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
2033             h->dequant8_coeff[1] = h->dequant8_buffer[0];
2034             break;
2035         }
2036
2037         for(q=0; q<52; q++){
2038             int shift = ff_div6[q];
2039             int idx = ff_rem6[q];
2040             for(x=0; x<64; x++)
2041                 h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
2042                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
2043                     h->pps.scaling_matrix8[i][x]) << shift;
2044         }
2045     }
2046 }
2047
2048 static void init_dequant4_coeff_table(H264Context *h){
2049     int i,j,q,x;
2050     const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
2051     for(i=0; i<6; i++ ){
2052         h->dequant4_coeff[i] = h->dequant4_buffer[i];
2053         for(j=0; j<i; j++){
2054             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
2055                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
2056                 break;
2057             }
2058         }
2059         if(j<i)
2060             continue;
2061
2062         for(q=0; q<52; q++){
2063             int shift = ff_div6[q] + 2;
2064             int idx = ff_rem6[q];
2065             for(x=0; x<16; x++)
2066                 h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
2067                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
2068                     h->pps.scaling_matrix4[i][x]) << shift;
2069         }
2070     }
2071 }
2072
2073 static void init_dequant_tables(H264Context *h){
2074     int i,x;
2075     init_dequant4_coeff_table(h);
2076     if(h->pps.transform_8x8_mode)
2077         init_dequant8_coeff_table(h);
2078     if(h->sps.transform_bypass){
2079         for(i=0; i<6; i++)
2080             for(x=0; x<16; x++)
2081                 h->dequant4_coeff[i][0][x] = 1<<6;
2082         if(h->pps.transform_8x8_mode)
2083             for(i=0; i<2; i++)
2084                 for(x=0; x<64; x++)
2085                     h->dequant8_coeff[i][0][x] = 1<<6;
2086     }
2087 }
2088
2089
2090 /**
2091  * allocates tables.
2092  * needs width/height
2093  */
2094 static int alloc_tables(H264Context *h){
2095     MpegEncContext * const s = &h->s;
2096     const int big_mb_num= s->mb_stride * (s->mb_height+1);
2097     int x,y;
2098
2099     CHECKED_ALLOCZ(h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t))
2100
2101     CHECKED_ALLOCZ(h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t))
2102     CHECKED_ALLOCZ(h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(uint8_t))
2103     CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))
2104
2105     CHECKED_ALLOCZ(h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t))
2106     CHECKED_ALLOCZ(h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t));
2107     CHECKED_ALLOCZ(h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t));
2108     CHECKED_ALLOCZ(h->direct_table, 32*big_mb_num * sizeof(uint8_t));
2109
2110     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(uint8_t));
2111     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
2112
2113     CHECKED_ALLOCZ(h->mb2b_xy  , big_mb_num * sizeof(uint32_t));
2114     CHECKED_ALLOCZ(h->mb2b8_xy , big_mb_num * sizeof(uint32_t));
2115     for(y=0; y<s->mb_height; y++){
2116         for(x=0; x<s->mb_width; x++){
2117             const int mb_xy= x + y*s->mb_stride;
2118             const int b_xy = 4*x + 4*y*h->b_stride;
2119             const int b8_xy= 2*x + 2*y*h->b8_stride;
2120
2121             h->mb2b_xy [mb_xy]= b_xy;
2122             h->mb2b8_xy[mb_xy]= b8_xy;
2123         }
2124     }
2125
2126     s->obmc_scratchpad = NULL;
2127
2128     if(!h->dequant4_coeff[0])
2129         init_dequant_tables(h);
2130
2131     return 0;
2132 fail:
2133     free_tables(h);
2134     return -1;
2135 }
2136
2137 /**
2138  * Mimic alloc_tables(), but for every context thread.
2139  */
2140 static void clone_tables(H264Context *dst, H264Context *src){
2141     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
2142     dst->non_zero_count           = src->non_zero_count;
2143     dst->slice_table              = src->slice_table;
2144     dst->cbp_table                = src->cbp_table;
2145     dst->mb2b_xy                  = src->mb2b_xy;
2146     dst->mb2b8_xy                 = src->mb2b8_xy;
2147     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
2148     dst->mvd_table[0]             = src->mvd_table[0];
2149     dst->mvd_table[1]             = src->mvd_table[1];
2150     dst->direct_table             = src->direct_table;
2151
2152     dst->s.obmc_scratchpad = NULL;
2153     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
2154 }
2155
2156 /**
2157  * Init context
2158  * Allocate buffers which are not shared amongst multiple threads.
2159  */
2160 static int context_init(H264Context *h){
2161     CHECKED_ALLOCZ(h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2162     CHECKED_ALLOCZ(h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2163
2164     return 0;
2165 fail:
2166     return -1; // free_tables will clean up for us
2167 }
2168
2169 static void common_init(H264Context *h){
2170     MpegEncContext * const s = &h->s;
2171
2172     s->width = s->avctx->width;
2173     s->height = s->avctx->height;
2174     s->codec_id= s->avctx->codec->id;
2175
2176     ff_h264_pred_init(&h->hpc, s->codec_id);
2177
2178     h->dequant_coeff_pps= -1;
2179     s->unrestricted_mv=1;
2180     s->decode=1; //FIXME
2181
2182     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
2183     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
2184 }
2185
2186 static int decode_init(AVCodecContext *avctx){
2187     H264Context *h= avctx->priv_data;
2188     MpegEncContext * const s = &h->s;
2189
2190     MPV_decode_defaults(s);
2191
2192     s->avctx = avctx;
2193     common_init(h);
2194
2195     s->out_format = FMT_H264;
2196     s->workaround_bugs= avctx->workaround_bugs;
2197
2198     // set defaults
2199 //    s->decode_mb= ff_h263_decode_mb;
2200     s->quarter_sample = 1;
2201     s->low_delay= 1;
2202     avctx->pix_fmt= PIX_FMT_YUV420P;
2203
2204     decode_init_vlc();
2205
2206     if(avctx->extradata_size > 0 && avctx->extradata &&
2207        *(char *)avctx->extradata == 1){
2208         h->is_avc = 1;
2209         h->got_avcC = 0;
2210     } else {
2211         h->is_avc = 0;
2212     }
2213
2214     h->thread_context[0] = h;
2215     return 0;
2216 }
2217
2218 static int frame_start(H264Context *h){
2219     MpegEncContext * const s = &h->s;
2220     int i;
2221
2222     if(MPV_frame_start(s, s->avctx) < 0)
2223         return -1;
2224     ff_er_frame_start(s);
2225     /*
2226      * MPV_frame_start uses pict_type to derive key_frame.
2227      * This is incorrect for H.264; IDR markings must be used.
2228      * Zero here; IDR markings per slice in frame or fields are OR'd in later.
2229      * See decode_nal_units().
2230      */
2231     s->current_picture_ptr->key_frame= 0;
2232
2233     assert(s->linesize && s->uvlinesize);
2234
2235     for(i=0; i<16; i++){
2236         h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
2237         h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
2238     }
2239     for(i=0; i<4; i++){
2240         h->block_offset[16+i]=
2241         h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2242         h->block_offset[24+16+i]=
2243         h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2244     }
2245
2246     /* can't be in alloc_tables because linesize isn't known there.
2247      * FIXME: redo bipred weight to not require extra buffer? */
2248     for(i = 0; i < s->avctx->thread_count; i++)
2249         if(!h->thread_context[i]->s.obmc_scratchpad)
2250             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
2251
2252     /* some macroblocks will be accessed before they're available */
2253     if(FRAME_MBAFF || s->avctx->thread_count > 1)
2254         memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(uint8_t));
2255
2256 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
2257     return 0;
2258 }
2259
2260 static inline void backup_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int simple){
2261     MpegEncContext * const s = &h->s;
2262     int i;
2263
2264     src_y  -=   linesize;
2265     src_cb -= uvlinesize;
2266     src_cr -= uvlinesize;
2267
2268     // There are two lines saved, the line above the the top macroblock of a pair,
2269     // and the line above the bottom macroblock
2270     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2271     for(i=1; i<17; i++){
2272         h->left_border[i]= src_y[15+i*  linesize];
2273     }
2274
2275     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
2276     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);
2277
2278     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2279         h->left_border[17  ]= h->top_borders[0][s->mb_x][16+7];
2280         h->left_border[17+9]= h->top_borders[0][s->mb_x][24+7];
2281         for(i=1; i<9; i++){
2282             h->left_border[i+17  ]= src_cb[7+i*uvlinesize];
2283             h->left_border[i+17+9]= src_cr[7+i*uvlinesize];
2284         }
2285         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
2286         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
2287     }
2288 }
2289
2290 static inline void xchg_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int xchg, int simple){
2291     MpegEncContext * const s = &h->s;
2292     int temp8, i;
2293     uint64_t temp64;
2294     int deblock_left;
2295     int deblock_top;
2296     int mb_xy;
2297
2298     if(h->deblocking_filter == 2) {
2299         mb_xy = s->mb_x + s->mb_y*s->mb_stride;
2300         deblock_left = h->slice_table[mb_xy] == h->slice_table[mb_xy - 1];
2301         deblock_top  = h->slice_table[mb_xy] == h->slice_table[h->top_mb_xy];
2302     } else {
2303         deblock_left = (s->mb_x > 0);
2304         deblock_top =  (s->mb_y > 0);
2305     }
2306
2307     src_y  -=   linesize + 1;
2308     src_cb -= uvlinesize + 1;
2309     src_cr -= uvlinesize + 1;
2310
2311 #define XCHG(a,b,t,xchg)\
2312 t= a;\
2313 if(xchg)\
2314     a= b;\
2315 b= t;
2316
2317     if(deblock_left){
2318         for(i = !deblock_top; i<17; i++){
2319             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2320         }
2321     }
2322
2323     if(deblock_top){
2324         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2325         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2326         if(s->mb_x+1 < s->mb_width){
2327             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2328         }
2329     }
2330
2331     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2332         if(deblock_left){
2333             for(i = !deblock_top; i<9; i++){
2334                 XCHG(h->left_border[i+17  ], src_cb[i*uvlinesize], temp8, xchg);
2335                 XCHG(h->left_border[i+17+9], src_cr[i*uvlinesize], temp8, xchg);
2336             }
2337         }
2338         if(deblock_top){
2339             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2340             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2341         }
2342     }
2343 }
2344
2345 static inline void backup_pair_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize){
2346     MpegEncContext * const s = &h->s;
2347     int i;
2348
2349     src_y  -= 2 *   linesize;
2350     src_cb -= 2 * uvlinesize;
2351     src_cr -= 2 * uvlinesize;
2352
2353     // There are two lines saved, the line above the the top macroblock of a pair,
2354     // and the line above the bottom macroblock
2355     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2356     h->left_border[1]= h->top_borders[1][s->mb_x][15];
2357     for(i=2; i<34; i++){
2358         h->left_border[i]= src_y[15+i*  linesize];
2359     }
2360
2361     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  32*linesize);
2362     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+32*linesize);
2363     *(uint64_t*)(h->top_borders[1][s->mb_x]+0)= *(uint64_t*)(src_y +  33*linesize);
2364     *(uint64_t*)(h->top_borders[1][s->mb_x]+8)= *(uint64_t*)(src_y +8+33*linesize);
2365
2366     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2367         h->left_border[34     ]= h->top_borders[0][s->mb_x][16+7];
2368         h->left_border[34+   1]= h->top_borders[1][s->mb_x][16+7];
2369         h->left_border[34+18  ]= h->top_borders[0][s->mb_x][24+7];
2370         h->left_border[34+18+1]= h->top_borders[1][s->mb_x][24+7];
2371         for(i=2; i<18; i++){
2372             h->left_border[i+34   ]= src_cb[7+i*uvlinesize];
2373             h->left_border[i+34+18]= src_cr[7+i*uvlinesize];
2374         }
2375         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+16*uvlinesize);
2376         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+16*uvlinesize);
2377         *(uint64_t*)(h->top_borders[1][s->mb_x]+16)= *(uint64_t*)(src_cb+17*uvlinesize);
2378         *(uint64_t*)(h->top_borders[1][s->mb_x]+24)= *(uint64_t*)(src_cr+17*uvlinesize);
2379     }
2380 }
2381
2382 static inline void xchg_pair_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int xchg){
2383     MpegEncContext * const s = &h->s;
2384     int temp8, i;
2385     uint64_t temp64;
2386     int deblock_left = (s->mb_x > 0);
2387     int deblock_top  = (s->mb_y > 1);
2388
2389     tprintf(s->avctx, "xchg_pair_border: src_y:%p src_cb:%p src_cr:%p ls:%d uvls:%d\n", src_y, src_cb, src_cr, linesize, uvlinesize);
2390
2391     src_y  -= 2 *   linesize + 1;
2392     src_cb -= 2 * uvlinesize + 1;
2393     src_cr -= 2 * uvlinesize + 1;
2394
2395 #define XCHG(a,b,t,xchg)\
2396 t= a;\
2397 if(xchg)\
2398     a= b;\
2399 b= t;
2400
2401     if(deblock_left){
2402         for(i = (!deblock_top)<<1; i<34; i++){
2403             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2404         }
2405     }
2406
2407     if(deblock_top){
2408         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2409         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2410         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+0), *(uint64_t*)(src_y +1 +linesize), temp64, xchg);
2411         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+8), *(uint64_t*)(src_y +9 +linesize), temp64, 1);
2412         if(s->mb_x+1 < s->mb_width){
2413             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2414             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x+1]), *(uint64_t*)(src_y +17 +linesize), temp64, 1);
2415         }
2416     }
2417
2418     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2419         if(deblock_left){
2420             for(i = (!deblock_top) << 1; i<18; i++){
2421                 XCHG(h->left_border[i+34   ], src_cb[i*uvlinesize], temp8, xchg);
2422                 XCHG(h->left_border[i+34+18], src_cr[i*uvlinesize], temp8, xchg);
2423             }
2424         }
2425         if(deblock_top){
2426             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2427             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2428             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+16), *(uint64_t*)(src_cb+1 +uvlinesize), temp64, 1);
2429             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+24), *(uint64_t*)(src_cr+1 +uvlinesize), temp64, 1);
2430         }
2431     }
2432 }
2433
2434 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
2435     MpegEncContext * const s = &h->s;
2436     const int mb_x= s->mb_x;
2437     const int mb_y= s->mb_y;
2438     const int mb_xy= mb_x + mb_y*s->mb_stride;
2439     const int mb_type= s->current_picture.mb_type[mb_xy];
2440     uint8_t  *dest_y, *dest_cb, *dest_cr;
2441     int linesize, uvlinesize /*dct_offset*/;
2442     int i;
2443     int *block_offset = &h->block_offset[0];
2444     const unsigned int bottom = mb_y & 1;
2445     const int transform_bypass = (s->qscale == 0 && h->sps.transform_bypass), is_h264 = (simple || s->codec_id == CODEC_ID_H264);
2446     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2447     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
2448
2449     dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2450     dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2451     dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2452
2453     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
2454     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
2455
2456     if (!simple && MB_FIELD) {
2457         linesize   = h->mb_linesize   = s->linesize * 2;
2458         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2459         block_offset = &h->block_offset[24];
2460         if(mb_y&1){ //FIXME move out of this func?
2461             dest_y -= s->linesize*15;
2462             dest_cb-= s->uvlinesize*7;
2463             dest_cr-= s->uvlinesize*7;
2464         }
2465         if(FRAME_MBAFF) {
2466             int list;
2467             for(list=0; list<h->list_count; list++){
2468                 if(!USES_LIST(mb_type, list))
2469                     continue;
2470                 if(IS_16X16(mb_type)){
2471                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2472                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2473                 }else{
2474                     for(i=0; i<16; i+=4){
2475                         //FIXME can refs be smaller than 8x8 when !direct_8x8_inference ?
2476                         int ref = h->ref_cache[list][scan8[i]];
2477                         if(ref >= 0)
2478                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2479                     }
2480                 }
2481             }
2482         }
2483     } else {
2484         linesize   = h->mb_linesize   = s->linesize;
2485         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2486 //        dct_offset = s->linesize * 16;
2487     }
2488
2489     if(transform_bypass){
2490         idct_dc_add =
2491         idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
2492     }else if(IS_8x8DCT(mb_type)){
2493         idct_dc_add = s->dsp.h264_idct8_dc_add;
2494         idct_add = s->dsp.h264_idct8_add;
2495     }else{
2496         idct_dc_add = s->dsp.h264_idct_dc_add;
2497         idct_add = s->dsp.h264_idct_add;
2498     }
2499
2500     if(!simple && FRAME_MBAFF && h->deblocking_filter && IS_INTRA(mb_type)
2501        && (!bottom || !IS_INTRA(s->current_picture.mb_type[mb_xy-s->mb_stride]))){
2502         int mbt_y = mb_y&~1;
2503         uint8_t *top_y  = s->current_picture.data[0] + (mbt_y * 16* s->linesize  ) + mb_x * 16;
2504         uint8_t *top_cb = s->current_picture.data[1] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2505         uint8_t *top_cr = s->current_picture.data[2] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2506         xchg_pair_border(h, top_y, top_cb, top_cr, s->linesize, s->uvlinesize, 1);
2507     }
2508
2509     if (!simple && IS_INTRA_PCM(mb_type)) {
2510         unsigned int x, y;
2511
2512         // The pixels are stored in h->mb array in the same order as levels,
2513         // copy them in output in the correct order.
2514         for(i=0; i<16; i++) {
2515             for (y=0; y<4; y++) {
2516                 for (x=0; x<4; x++) {
2517                     *(dest_y + block_offset[i] + y*linesize + x) = h->mb[i*16+y*4+x];
2518                 }
2519             }
2520         }
2521         for(i=16; i<16+4; i++) {
2522             for (y=0; y<4; y++) {
2523                 for (x=0; x<4; x++) {
2524                     *(dest_cb + block_offset[i] + y*uvlinesize + x) = h->mb[i*16+y*4+x];
2525                 }
2526             }
2527         }
2528         for(i=20; i<20+4; i++) {
2529             for (y=0; y<4; y++) {
2530                 for (x=0; x<4; x++) {
2531                     *(dest_cr + block_offset[i] + y*uvlinesize + x) = h->mb[i*16+y*4+x];
2532                 }
2533             }
2534         }
2535     } else {
2536         if(IS_INTRA(mb_type)){
2537             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2538                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
2539
2540             if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2541                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2542                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2543             }
2544
2545             if(IS_INTRA4x4(mb_type)){
2546                 if(simple || !s->encoding){
2547                     if(IS_8x8DCT(mb_type)){
2548                         for(i=0; i<16; i+=4){
2549                             uint8_t * const ptr= dest_y + block_offset[i];
2550                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2551                             const int nnz = h->non_zero_count_cache[ scan8[i] ];
2552                             h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
2553                                                    (h->topright_samples_available<<i)&0x4000, linesize);
2554                             if(nnz){
2555                                 if(nnz == 1 && h->mb[i*16])
2556                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2557                                 else
2558                                     idct_add(ptr, h->mb + i*16, linesize);
2559                             }
2560                         }
2561                     }else
2562                     for(i=0; i<16; i++){
2563                         uint8_t * const ptr= dest_y + block_offset[i];
2564                         uint8_t *topright;
2565                         const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2566                         int nnz, tr;
2567
2568                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
2569                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
2570                             assert(mb_y || linesize <= block_offset[i]);
2571                             if(!topright_avail){
2572                                 tr= ptr[3 - linesize]*0x01010101;
2573                                 topright= (uint8_t*) &tr;
2574                             }else
2575                                 topright= ptr + 4 - linesize;
2576                         }else
2577                             topright= NULL;
2578
2579                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
2580                         nnz = h->non_zero_count_cache[ scan8[i] ];
2581                         if(nnz){
2582                             if(is_h264){
2583                                 if(nnz == 1 && h->mb[i*16])
2584                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2585                                 else
2586                                     idct_add(ptr, h->mb + i*16, linesize);
2587                             }else
2588                                 svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
2589                         }
2590                     }
2591                 }
2592             }else{
2593                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
2594                 if(is_h264){
2595                     if(!transform_bypass)
2596                         h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[0][s->qscale][0]);
2597                 }else
2598                     svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
2599             }
2600             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2601                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
2602         }else if(is_h264){
2603             hl_motion(h, dest_y, dest_cb, dest_cr,
2604                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2605                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2606                       s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
2607         }
2608
2609
2610         if(!IS_INTRA4x4(mb_type)){
2611             if(is_h264){
2612                 if(IS_INTRA16x16(mb_type)){
2613                     for(i=0; i<16; i++){
2614                         if(h->non_zero_count_cache[ scan8[i] ])
2615                             idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2616                         else if(h->mb[i*16])
2617                             idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2618                     }
2619                 }else{
2620                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2621                     for(i=0; i<16; i+=di){
2622                         int nnz = h->non_zero_count_cache[ scan8[i] ];
2623                         if(nnz){
2624                             if(nnz==1 && h->mb[i*16])
2625                                 idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2626                             else
2627                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2628                         }
2629                     }
2630                 }
2631             }else{
2632                 for(i=0; i<16; i++){
2633                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
2634                         uint8_t * const ptr= dest_y + block_offset[i];
2635                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2636                     }
2637                 }
2638             }
2639         }
2640
2641         if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2642             uint8_t *dest[2] = {dest_cb, dest_cr};
2643             if(transform_bypass){
2644                 idct_add = idct_dc_add = s->dsp.add_pixels4;
2645             }else{
2646                 idct_add = s->dsp.h264_idct_add;
2647                 idct_dc_add = s->dsp.h264_idct_dc_add;
2648                 chroma_dc_dequant_idct_c(h->mb + 16*16, h->chroma_qp[0], h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp[0]][0]);
2649                 chroma_dc_dequant_idct_c(h->mb + 16*16+4*16, h->chroma_qp[1], h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp[1]][0]);
2650             }
2651             if(is_h264){
2652                 for(i=16; i<16+8; i++){
2653                     if(h->non_zero_count_cache[ scan8[i] ])
2654                         idct_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2655                     else if(h->mb[i*16])
2656                         idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2657                 }
2658             }else{
2659                 for(i=16; i<16+8; i++){
2660                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2661                         uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
2662                         svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, chroma_qp[s->qscale + 12] - 12, 2);
2663                     }
2664                 }
2665             }
2666         }
2667     }
2668     if(h->deblocking_filter) {
2669         if (!simple && FRAME_MBAFF) {
2670             //FIXME try deblocking one mb at a time?
2671             // the reduction in load/storing mvs and such might outweigh the extra backup/xchg_border
2672             const int mb_y = s->mb_y - 1;
2673             uint8_t  *pair_dest_y, *pair_dest_cb, *pair_dest_cr;
2674             const int mb_xy= mb_x + mb_y*s->mb_stride;
2675             const int mb_type_top   = s->current_picture.mb_type[mb_xy];
2676             const int mb_type_bottom= s->current_picture.mb_type[mb_xy+s->mb_stride];
2677             if (!bottom) return;
2678             pair_dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2679             pair_dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2680             pair_dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2681
2682             if(IS_INTRA(mb_type_top | mb_type_bottom))
2683                 xchg_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize, 0);
2684
2685             backup_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize);
2686             // deblock a pair
2687             // top
2688             s->mb_y--;
2689             tprintf(h->s.avctx, "call mbaff filter_mb mb_x:%d mb_y:%d pair_dest_y = %p, dest_y = %p\n", mb_x, mb_y, pair_dest_y, dest_y);
2690             fill_caches(h, mb_type_top, 1); //FIXME don't fill stuff which isn't used by filter_mb
2691             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2692             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2693             filter_mb(h, mb_x, mb_y, pair_dest_y, pair_dest_cb, pair_dest_cr, linesize, uvlinesize);
2694             // bottom
2695             s->mb_y++;
2696             tprintf(h->s.avctx, "call mbaff filter_mb\n");
2697             fill_caches(h, mb_type_bottom, 1); //FIXME don't fill stuff which isn't used by filter_mb
2698             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2699             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2700             filter_mb(h, mb_x, mb_y+1, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2701         } else {
2702             tprintf(h->s.avctx, "call filter_mb\n");
2703             backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
2704             fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
2705             filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2706         }
2707     }
2708 }
2709
2710 /**
2711  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2712  */
2713 static void hl_decode_mb_simple(H264Context *h){
2714     hl_decode_mb_internal(h, 1);
2715 }
2716
2717 /**
2718  * Process a macroblock; this handles edge cases, such as interlacing.
2719  */
2720 static void av_noinline hl_decode_mb_complex(H264Context *h){
2721     hl_decode_mb_internal(h, 0);
2722 }
2723
2724 static void hl_decode_mb(H264Context *h){
2725     MpegEncContext * const s = &h->s;
2726     const int mb_x= s->mb_x;
2727     const int mb_y= s->mb_y;
2728     const int mb_xy= mb_x + mb_y*s->mb_stride;
2729     const int mb_type= s->current_picture.mb_type[mb_xy];
2730     int is_complex = FRAME_MBAFF || MB_FIELD || IS_INTRA_PCM(mb_type) || s->codec_id != CODEC_ID_H264 || (ENABLE_GRAY && (s->flags&CODEC_FLAG_GRAY)) || s->encoding;
2731
2732     if(!s->decode)
2733         return;
2734
2735     if (is_complex)
2736         hl_decode_mb_complex(h);
2737     else hl_decode_mb_simple(h);
2738 }
2739
2740 static void pic_as_field(Picture *pic, const int parity){
2741     int i;
2742     for (i = 0; i < 4; ++i) {
2743         if (parity == PICT_BOTTOM_FIELD)
2744             pic->data[i] += pic->linesize[i];
2745         pic->reference = parity;
2746         pic->linesize[i] *= 2;
2747     }
2748 }
2749
2750 static int split_field_copy(Picture *dest, Picture *src,
2751                             int parity, int id_add){
2752     int match = !!(src->reference & parity);
2753
2754     if (match) {
2755         *dest = *src;
2756         pic_as_field(dest, parity);
2757         dest->pic_id *= 2;
2758         dest->pic_id += id_add;
2759     }
2760
2761     return match;
2762 }
2763
2764 /**
2765  * Split one reference list into field parts, interleaving by parity
2766  * as per H.264 spec section 8.2.4.2.5. Output fields have their data pointers
2767  * set to look at the actual start of data for that field.
2768  *
2769  * @param dest output list
2770  * @param dest_len maximum number of fields to put in dest
2771  * @param src the source reference list containing fields and/or field pairs
2772  *            (aka short_ref/long_ref, or
2773  *             refFrameListXShortTerm/refFrameListLongTerm in spec-speak)
2774  * @param src_len number of Picture's in source (pairs and unmatched fields)
2775  * @param parity the parity of the picture being decoded/needing
2776  *        these ref pics (PICT_{TOP,BOTTOM}_FIELD)
2777  * @return number of fields placed in dest
2778  */
2779 static int split_field_half_ref_list(Picture *dest, int dest_len,
2780                                      Picture *src,  int src_len,  int parity){
2781     int same_parity   = 1;
2782     int same_i        = 0;
2783     int opp_i         = 0;
2784     int out_i;
2785     int field_output;
2786
2787     for (out_i = 0; out_i < dest_len; out_i += field_output) {
2788         if (same_parity && same_i < src_len) {
2789             field_output = split_field_copy(dest + out_i, src + same_i,
2790                                             parity, 1);
2791             same_parity = !field_output;
2792             same_i++;
2793
2794         } else if (opp_i < src_len) {
2795             field_output = split_field_copy(dest + out_i, src + opp_i,
2796                                             PICT_FRAME - parity, 0);
2797             same_parity = field_output;
2798             opp_i++;
2799
2800         } else {
2801             break;
2802         }
2803     }
2804
2805     return out_i;
2806 }
2807
2808 /**
2809  * Split the reference frame list into a reference field list.
2810  * This implements H.264 spec 8.2.4.2.5 for a combined input list.
2811  * The input list contains both reference field pairs and
2812  * unmatched reference fields; it is ordered as spec describes
2813  * RefPicListX for frames in 8.2.4.2.1 and 8.2.4.2.3, except that
2814  * unmatched field pairs are also present. Conceptually this is equivalent
2815  * to concatenation of refFrameListXShortTerm with refFrameListLongTerm.
2816  *
2817  * @param dest output reference list where ordered fields are to be placed
2818  * @param dest_len max number of fields to place at dest
2819  * @param src source reference list, as described above
2820  * @param src_len number of pictures (pairs and unmatched fields) in src
2821  * @param parity parity of field being currently decoded
2822  *        (one of PICT_{TOP,BOTTOM}_FIELD)
2823  * @param long_i index into src array that holds first long reference picture,
2824  *        or src_len if no long refs present.
2825  */
2826 static int split_field_ref_list(Picture *dest, int dest_len,
2827                                 Picture *src,  int src_len,
2828                                 int parity,    int long_i){
2829
2830     int i = split_field_half_ref_list(dest, dest_len, src, long_i, parity);
2831     dest += i;
2832     dest_len -= i;
2833
2834     i += split_field_half_ref_list(dest, dest_len, src + long_i,
2835                                    src_len - long_i, parity);
2836     return i;
2837 }
2838
2839 /**
2840  * fills the default_ref_list.
2841  */
2842 static int fill_default_ref_list(H264Context *h){
2843     MpegEncContext * const s = &h->s;
2844     int i;
2845     int smallest_poc_greater_than_current = -1;
2846     int structure_sel;
2847     Picture sorted_short_ref[32];
2848     Picture field_entry_list[2][32];
2849     Picture *frame_list[2];
2850
2851     if (FIELD_PICTURE) {
2852         structure_sel = PICT_FRAME;
2853         frame_list[0] = field_entry_list[0];
2854         frame_list[1] = field_entry_list[1];
2855     } else {
2856         structure_sel = 0;
2857         frame_list[0] = h->default_ref_list[0];
2858         frame_list[1] = h->default_ref_list[1];
2859     }
2860
2861     if(h->slice_type==FF_B_TYPE){
2862         int list;
2863         int len[2];
2864         int short_len[2];
2865         int out_i;
2866         int limit= INT_MIN;
2867
2868         /* sort frame according to poc in B slice */
2869         for(out_i=0; out_i<h->short_ref_count; out_i++){
2870             int best_i=INT_MIN;
2871             int best_poc=INT_MAX;
2872
2873             for(i=0; i<h->short_ref_count; i++){
2874                 const int poc= h->short_ref[i]->poc;
2875                 if(poc > limit && poc < best_poc){
2876                     best_poc= poc;
2877                     best_i= i;
2878                 }
2879             }
2880
2881             assert(best_i != INT_MIN);
2882
2883             limit= best_poc;
2884             sorted_short_ref[out_i]= *h->short_ref[best_i];
2885             tprintf(h->s.avctx, "sorted poc: %d->%d poc:%d fn:%d\n", best_i, out_i, sorted_short_ref[out_i].poc, sorted_short_ref[out_i].frame_num);
2886             if (-1 == smallest_poc_greater_than_current) {
2887                 if (h->short_ref[best_i]->poc >= s->current_picture_ptr->poc) {
2888                     smallest_poc_greater_than_current = out_i;
2889                 }
2890             }
2891         }
2892
2893         tprintf(h->s.avctx, "current poc: %d, smallest_poc_greater_than_current: %d\n", s->current_picture_ptr->poc, smallest_poc_greater_than_current);
2894
2895         // find the largest poc
2896         for(list=0; list<2; list++){
2897             int index = 0;
2898             int j= -99;
2899             int step= list ? -1 : 1;
2900
2901             for(i=0; i<h->short_ref_count && index < h->ref_count[list]; i++, j+=step) {
2902                 int sel;
2903                 while(j<0 || j>= h->short_ref_count){
2904                     if(j != -99 && step == (list ? -1 : 1))
2905                         return -1;
2906                     step = -step;
2907                     j= smallest_poc_greater_than_current + (step>>1);
2908                 }
2909                 sel = sorted_short_ref[j].reference | structure_sel;
2910                 if(sel != PICT_FRAME) continue;
2911                 frame_list[list][index  ]= sorted_short_ref[j];
2912                 frame_list[list][index++].pic_id= sorted_short_ref[j].frame_num;
2913             }
2914             short_len[list] = index;
2915
2916             for(i = 0; i < 16 && index < h->ref_count[ list ]; i++){
2917                 int sel;
2918                 if(h->long_ref[i] == NULL) continue;
2919                 sel = h->long_ref[i]->reference | structure_sel;
2920                 if(sel != PICT_FRAME) continue;
2921
2922                 frame_list[ list ][index  ]= *h->long_ref[i];
2923                 frame_list[ list ][index++].pic_id= i;
2924             }
2925             len[list] = index;
2926         }
2927
2928         for(list=0; list<2; list++){
2929             if (FIELD_PICTURE)
2930                 len[list] = split_field_ref_list(h->default_ref_list[list],
2931                                                  h->ref_count[list],
2932                                                  frame_list[list],
2933                                                  len[list],
2934                                                  s->picture_structure,
2935                                                  short_len[list]);
2936
2937             // swap the two first elements of L1 when L0 and L1 are identical
2938             if(list && len[0] > 1 && len[0] == len[1])
2939                 for(i=0; h->default_ref_list[0][i].data[0] == h->default_ref_list[1][i].data[0]; i++)
2940                     if(i == len[0]){
2941                         FFSWAP(Picture, h->default_ref_list[1][0], h->default_ref_list[1][1]);
2942                         break;
2943                     }
2944
2945             if(len[list] < h->ref_count[ list ])
2946                 memset(&h->default_ref_list[list][len[list]], 0, sizeof(Picture)*(h->ref_count[ list ] - len[list]));
2947         }
2948
2949
2950     }else{
2951         int index=0;
2952         int short_len;
2953         for(i=0; i<h->short_ref_count; i++){
2954             int sel;
2955             sel = h->short_ref[i]->reference | structure_sel;
2956             if(sel != PICT_FRAME) continue;
2957             frame_list[0][index  ]= *h->short_ref[i];
2958             frame_list[0][index++].pic_id= h->short_ref[i]->frame_num;
2959         }
2960         short_len = index;
2961         for(i = 0; i < 16; i++){
2962             int sel;
2963             if(h->long_ref[i] == NULL) continue;
2964             sel = h->long_ref[i]->reference | structure_sel;
2965             if(sel != PICT_FRAME) continue;
2966             frame_list[0][index  ]= *h->long_ref[i];
2967             frame_list[0][index++].pic_id= i;
2968         }
2969
2970         if (FIELD_PICTURE)
2971             index = split_field_ref_list(h->default_ref_list[0],
2972                                          h->ref_count[0], frame_list[0],
2973                                          index, s->picture_structure,
2974                                          short_len);
2975
2976         if(index < h->ref_count[0])
2977             memset(&h->default_ref_list[0][index], 0, sizeof(Picture)*(h->ref_count[0] - index));
2978     }
2979 #ifdef TRACE
2980     for (i=0; i<h->ref_count[0]; i++) {
2981         tprintf(h->s.avctx, "List0: %s fn:%d 0x%p\n", (h->default_ref_list[0][i].long_ref ? "LT" : "ST"), h->default_ref_list[0][i].pic_id, h->default_ref_list[0][i].data[0]);
2982     }
2983     if(h->slice_type==FF_B_TYPE){
2984         for (i=0; i<h->ref_count[1]; i++) {
2985             tprintf(h->s.avctx, "List1: %s fn:%d 0x%p\n", (h->default_ref_list[1][i].long_ref ? "LT" : "ST"), h->default_ref_list[1][i].pic_id, h->default_ref_list[1][i].data[0]);
2986         }
2987     }
2988 #endif
2989     return 0;
2990 }
2991
2992 static void print_short_term(H264Context *h);
2993 static void print_long_term(H264Context *h);
2994
2995 /**
2996  * Extract structure information about the picture described by pic_num in
2997  * the current decoding context (frame or field). Note that pic_num is
2998  * picture number without wrapping (so, 0<=pic_num<max_pic_num).
2999  * @param pic_num picture number for which to extract structure information
3000  * @param structure one of PICT_XXX describing structure of picture
3001  *                      with pic_num
3002  * @return frame number (short term) or long term index of picture
3003  *         described by pic_num
3004  */
3005 static int pic_num_extract(H264Context *h, int pic_num, int *structure){
3006     MpegEncContext * const s = &h->s;
3007
3008     *structure = s->picture_structure;
3009     if(FIELD_PICTURE){
3010         if (!(pic_num & 1))
3011             /* opposite field */
3012             *structure ^= PICT_FRAME;
3013         pic_num >>= 1;
3014     }
3015
3016     return pic_num;
3017 }
3018
3019 static int decode_ref_pic_list_reordering(H264Context *h){
3020     MpegEncContext * const s = &h->s;
3021     int list, index, pic_structure;
3022
3023     print_short_term(h);
3024     print_long_term(h);
3025     if(h->slice_type==FF_I_TYPE || h->slice_type==FF_SI_TYPE) return 0; //FIXME move before func
3026
3027     for(list=0; list<h->list_count; list++){
3028         memcpy(h->ref_list[list], h->default_ref_list[list], sizeof(Picture)*h->ref_count[list]);
3029
3030         if(get_bits1(&s->gb)){
3031             int pred= h->curr_pic_num;
3032
3033             for(index=0; ; index++){
3034                 unsigned int reordering_of_pic_nums_idc= get_ue_golomb(&s->gb);
3035                 unsigned int pic_id;
3036                 int i;
3037                 Picture *ref = NULL;
3038
3039                 if(reordering_of_pic_nums_idc==3)
3040                     break;
3041
3042                 if(index >= h->ref_count[list]){
3043                     av_log(h->s.avctx, AV_LOG_ERROR, "reference count overflow\n");
3044                     return -1;
3045                 }
3046
3047                 if(reordering_of_pic_nums_idc<3){
3048                     if(reordering_of_pic_nums_idc<2){
3049                         const unsigned int abs_diff_pic_num= get_ue_golomb(&s->gb) + 1;
3050                         int frame_num;
3051
3052                         if(abs_diff_pic_num > h->max_pic_num){
3053                             av_log(h->s.avctx, AV_LOG_ERROR, "abs_diff_pic_num overflow\n");
3054                             return -1;
3055                         }
3056
3057                         if(reordering_of_pic_nums_idc == 0) pred-= abs_diff_pic_num;
3058                         else                                pred+= abs_diff_pic_num;
3059                         pred &= h->max_pic_num - 1;
3060
3061                         frame_num = pic_num_extract(h, pred, &pic_structure);
3062
3063                         for(i= h->short_ref_count-1; i>=0; i--){
3064                             ref = h->short_ref[i];
3065                             assert(ref->reference);
3066                             assert(!ref->long_ref);
3067                             if(ref->data[0] != NULL &&
3068                                    ref->frame_num == frame_num &&
3069                                    (ref->reference & pic_structure) &&
3070                                    ref->long_ref == 0) // ignore non existing pictures by testing data[0] pointer
3071                                 break;
3072                         }
3073                         if(i>=0)
3074                             ref->pic_id= pred;
3075                     }else{
3076                         int long_idx;
3077                         pic_id= get_ue_golomb(&s->gb); //long_term_pic_idx
3078
3079                         long_idx= pic_num_extract(h, pic_id, &pic_structure);
3080
3081                         if(long_idx>31){
3082                             av_log(h->s.avctx, AV_LOG_ERROR, "long_term_pic_idx overflow\n");
3083                             return -1;
3084                         }
3085                         ref = h->long_ref[long_idx];
3086                         assert(!(ref && !ref->reference));
3087                         if(ref && (ref->reference & pic_structure)){
3088                             ref->pic_id= pic_id;
3089                             assert(ref->long_ref);
3090                             i=0;
3091                         }else{
3092                             i=-1;
3093                         }
3094                     }
3095
3096                     if (i < 0) {
3097                         av_log(h->s.avctx, AV_LOG_ERROR, "reference picture missing during reorder\n");
3098                         memset(&h->ref_list[list][index], 0, sizeof(Picture)); //FIXME
3099                     } else {
3100                         for(i=index; i+1<h->ref_count[list]; i++){
3101                             if(ref->long_ref == h->ref_list[list][i].long_ref && ref->pic_id == h->ref_list[list][i].pic_id)
3102                                 break;
3103                         }
3104                         for(; i > index; i--){
3105                             h->ref_list[list][i]= h->ref_list[list][i-1];
3106                         }
3107                         h->ref_list[list][index]= *ref;
3108                         if (FIELD_PICTURE){
3109                             pic_as_field(&h->ref_list[list][index], pic_structure);
3110                         }
3111                     }
3112                 }else{
3113                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal reordering_of_pic_nums_idc\n");
3114                     return -1;
3115                 }
3116             }
3117         }
3118     }
3119     for(list=0; list<h->list_count; list++){
3120         for(index= 0; index < h->ref_count[list]; index++){
3121             if(!h->ref_list[list][index].data[0])
3122                 h->ref_list[list][index]= s->current_picture;
3123         }
3124     }
3125
3126     if(h->slice_type==FF_B_TYPE && !h->direct_spatial_mv_pred)
3127         direct_dist_scale_factor(h);
3128     direct_ref_list_init(h);
3129     return 0;
3130 }
3131
3132 static void fill_mbaff_ref_list(H264Context *h){
3133     int list, i, j;
3134     for(list=0; list<2; list++){ //FIXME try list_count
3135         for(i=0; i<h->ref_count[list]; i++){
3136             Picture *frame = &h->ref_list[list][i];
3137             Picture *field = &h->ref_list[list][16+2*i];
3138             field[0] = *frame;
3139             for(j=0; j<3; j++)
3140                 field[0].linesize[j] <<= 1;
3141             field[0].reference = PICT_TOP_FIELD;
3142             field[1] = field[0];
3143             for(j=0; j<3; j++)
3144                 field[1].data[j] += frame->linesize[j];
3145             field[1].reference = PICT_BOTTOM_FIELD;
3146
3147             h->luma_weight[list][16+2*i] = h->luma_weight[list][16+2*i+1] = h->luma_weight[list][i];
3148             h->luma_offset[list][16+2*i] = h->luma_offset[list][16+2*i+1] = h->luma_offset[list][i];
3149             for(j=0; j<2; j++){
3150                 h->chroma_weight[list][16+2*i][j] = h->chroma_weight[list][16+2*i+1][j] = h->chroma_weight[list][i][j];
3151                 h->chroma_offset[list][16+2*i][j] = h->chroma_offset[list][16+2*i+1][j] = h->chroma_offset[list][i][j];
3152             }
3153         }
3154     }
3155     for(j=0; j<h->ref_count[1]; j++){
3156         for(i=0; i<h->ref_count[0]; i++)
3157             h->implicit_weight[j][16+2*i] = h->implicit_weight[j][16+2*i+1] = h->implicit_weight[j][i];
3158         memcpy(h->implicit_weight[16+2*j],   h->implicit_weight[j], sizeof(*h->implicit_weight));
3159         memcpy(h->implicit_weight[16+2*j+1], h->implicit_weight[j], sizeof(*h->implicit_weight));
3160     }
3161 }
3162
3163 static int pred_weight_table(H264Context *h){
3164     MpegEncContext * const s = &h->s;
3165     int list, i;
3166     int luma_def, chroma_def;
3167
3168     h->use_weight= 0;
3169     h->use_weight_chroma= 0;
3170     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
3171     h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
3172     luma_def = 1<<h->luma_log2_weight_denom;
3173     chroma_def = 1<<h->chroma_log2_weight_denom;
3174
3175     for(list=0; list<2; list++){
3176         for(i=0; i<h->ref_count[list]; i++){
3177             int luma_weight_flag, chroma_weight_flag;
3178
3179             luma_weight_flag= get_bits1(&s->gb);
3180             if(luma_weight_flag){
3181                 h->luma_weight[list][i]= get_se_golomb(&s->gb);
3182                 h->luma_offset[list][i]= get_se_golomb(&s->gb);
3183                 if(   h->luma_weight[list][i] != luma_def
3184                    || h->luma_offset[list][i] != 0)
3185                     h->use_weight= 1;
3186             }else{
3187                 h->luma_weight[list][i]= luma_def;
3188                 h->luma_offset[list][i]= 0;
3189             }
3190
3191             chroma_weight_flag= get_bits1(&s->gb);
3192             if(chroma_weight_flag){
3193                 int j;
3194                 for(j=0; j<2; j++){
3195                     h->chroma_weight[list][i][j]= get_se_golomb(&s->gb);
3196                     h->chroma_offset[list][i][j]= get_se_golomb(&s->gb);
3197                     if(   h->chroma_weight[list][i][j] != chroma_def
3198                        || h->chroma_offset[list][i][j] != 0)
3199                         h->use_weight_chroma= 1;
3200                 }
3201             }else{
3202                 int j;
3203                 for(j=0; j<2; j++){
3204                     h->chroma_weight[list][i][j]= chroma_def;
3205                     h->chroma_offset[list][i][j]= 0;
3206                 }
3207             }
3208         }
3209         if(h->slice_type != FF_B_TYPE) break;
3210     }
3211     h->use_weight= h->use_weight || h->use_weight_chroma;
3212     return 0;
3213 }
3214
3215 static void implicit_weight_table(H264Context *h){
3216     MpegEncContext * const s = &h->s;
3217     int ref0, ref1;
3218     int cur_poc = s->current_picture_ptr->poc;
3219
3220     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1
3221        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
3222         h->use_weight= 0;
3223         h->use_weight_chroma= 0;
3224         return;
3225     }
3226
3227     h->use_weight= 2;
3228     h->use_weight_chroma= 2;
3229     h->luma_log2_weight_denom= 5;
3230     h->chroma_log2_weight_denom= 5;
3231
3232     for(ref0=0; ref0 < h->ref_count[0]; ref0++){
3233         int poc0 = h->ref_list[0][ref0].poc;
3234         for(ref1=0; ref1 < h->ref_count[1]; ref1++){
3235             int poc1 = h->ref_list[1][ref1].poc;
3236             int td = av_clip(poc1 - poc0, -128, 127);
3237             if(td){
3238                 int tb = av_clip(cur_poc - poc0, -128, 127);
3239                 int tx = (16384 + (FFABS(td) >> 1)) / td;
3240                 int dist_scale_factor = av_clip((tb*tx + 32) >> 6, -1024, 1023) >> 2;
3241                 if(dist_scale_factor < -64 || dist_scale_factor > 128)
3242                     h->implicit_weight[ref0][ref1] = 32;
3243                 else
3244                     h->implicit_weight[ref0][ref1] = 64 - dist_scale_factor;
3245             }else
3246                 h->implicit_weight[ref0][ref1] = 32;
3247         }
3248     }
3249 }
3250
3251 /**
3252  * Mark a picture as no longer needed for reference. The refmask
3253  * argument allows unreferencing of individual fields or the whole frame.
3254  * If the picture becomes entirely unreferenced, but is being held for
3255  * display purposes, it is marked as such.
3256  * @param refmask mask of fields to unreference; the mask is bitwise
3257  *                anded with the reference marking of pic
3258  * @return non-zero if pic becomes entirely unreferenced (except possibly
3259  *         for display purposes) zero if one of the fields remains in
3260  *         reference
3261  */
3262 static inline int unreference_pic(H264Context *h, Picture *pic, int refmask){
3263     int i;
3264     if (pic->reference &= refmask) {
3265         return 0;
3266     } else {
3267         if(pic == h->delayed_output_pic)
3268             pic->reference=DELAYED_PIC_REF;
3269         else{
3270             for(i = 0; h->delayed_pic[i]; i++)
3271                 if(pic == h->delayed_pic[i]){
3272                     pic->reference=DELAYED_PIC_REF;
3273                     break;
3274                 }
3275         }
3276         return 1;
3277     }
3278 }
3279
3280 /**
3281  * instantaneous decoder refresh.
3282  */
3283 static void idr(H264Context *h){
3284     int i;
3285
3286     for(i=0; i<16; i++){
3287         if (h->long_ref[i] != NULL) {
3288             unreference_pic(h, h->long_ref[i], 0);
3289             h->long_ref[i]= NULL;
3290         }
3291     }
3292     h->long_ref_count=0;
3293
3294     for(i=0; i<h->short_ref_count; i++){
3295         unreference_pic(h, h->short_ref[i], 0);
3296         h->short_ref[i]= NULL;
3297     }
3298     h->short_ref_count=0;
3299 }
3300
3301 /* forget old pics after a seek */
3302 static void flush_dpb(AVCodecContext *avctx){
3303     H264Context *h= avctx->priv_data;
3304     int i;
3305     for(i=0; i<16; i++) {
3306         if(h->delayed_pic[i])
3307             h->delayed_pic[i]->reference= 0;
3308         h->delayed_pic[i]= NULL;
3309     }
3310     if(h->delayed_output_pic)
3311         h->delayed_output_pic->reference= 0;
3312     h->delayed_output_pic= NULL;
3313     idr(h);
3314     if(h->s.current_picture_ptr)
3315         h->s.current_picture_ptr->reference= 0;
3316     h->s.first_field= 0;
3317     ff_mpeg_flush(avctx);
3318 }
3319
3320 /**
3321  * Find a Picture in the short term reference list by frame number.
3322  * @param frame_num frame number to search for
3323  * @param idx the index into h->short_ref where returned picture is found
3324  *            undefined if no picture found.
3325  * @return pointer to the found picture, or NULL if no pic with the provided
3326  *                 frame number is found
3327  */
3328 static Picture * find_short(H264Context *h, int frame_num, int *idx){
3329     MpegEncContext * const s = &h->s;
3330     int i;
3331
3332     for(i=0; i<h->short_ref_count; i++){
3333         Picture *pic= h->short_ref[i];
3334         if(s->avctx->debug&FF_DEBUG_MMCO)
3335             av_log(h->s.avctx, AV_LOG_DEBUG, "%d %d %p\n", i, pic->frame_num, pic);
3336         if(pic->frame_num == frame_num) {
3337             *idx = i;
3338             return pic;
3339         }
3340     }
3341     return NULL;
3342 }
3343
3344 /**
3345  * Remove a picture from the short term reference list by its index in
3346  * that list.  This does no checking on the provided index; it is assumed
3347  * to be valid. Other list entries are shifted down.
3348  * @param i index into h->short_ref of picture to remove.
3349  */
3350 static void remove_short_at_index(H264Context *h, int i){
3351     assert(i > 0 && i < h->short_ref_count);
3352     h->short_ref[i]= NULL;
3353     if (--h->short_ref_count)
3354         memmove(&h->short_ref[i], &h->short_ref[i+1], (h->short_ref_count - i)*sizeof(Picture*));
3355 }
3356
3357 /**
3358  *
3359  * @return the removed picture or NULL if an error occurs
3360  */
3361 static Picture * remove_short(H264Context *h, int frame_num){
3362     MpegEncContext * const s = &h->s;
3363     Picture *pic;
3364     int i;
3365
3366     if(s->avctx->debug&FF_DEBUG_MMCO)
3367         av_log(h->s.avctx, AV_LOG_DEBUG, "remove short %d count %d\n", frame_num, h->short_ref_count);
3368
3369     pic = find_short(h, frame_num, &i);
3370     if (pic)
3371         remove_short_at_index(h, i);
3372
3373     return pic;
3374 }
3375
3376 /**
3377  * Remove a picture from the long term reference list by its index in
3378  * that list.  This does no checking on the provided index; it is assumed
3379  * to be valid. The removed entry is set to NULL. Other entries are unaffected.
3380  * @param i index into h->long_ref of picture to remove.
3381  */
3382 static void remove_long_at_index(H264Context *h, int i){
3383     h->long_ref[i]= NULL;
3384     h->long_ref_count--;
3385 }
3386
3387 /**
3388  *
3389  * @return the removed picture or NULL if an error occurs
3390  */
3391 static Picture * remove_long(H264Context *h, int i){
3392     Picture *pic;
3393
3394     pic= h->long_ref[i];
3395     if (pic)
3396         remove_long_at_index(h, i);
3397
3398     return pic;
3399 }
3400
3401 /**
3402  * print short term list
3403  */
3404 static void print_short_term(H264Context *h) {
3405     uint32_t i;
3406     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3407         av_log(h->s.avctx, AV_LOG_DEBUG, "short term list:\n");
3408         for(i=0; i<h->short_ref_count; i++){
3409             Picture *pic= h->short_ref[i];
3410             av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3411         }
3412     }
3413 }
3414
3415 /**
3416  * print long term list
3417  */
3418 static void print_long_term(H264Context *h) {
3419     uint32_t i;
3420     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3421         av_log(h->s.avctx, AV_LOG_DEBUG, "long term list:\n");
3422         for(i = 0; i < 16; i++){
3423             Picture *pic= h->long_ref[i];
3424             if&